Merge #4649

4649: Don't store the vectors in the documents database r=dureuill a=irevoire

# Pull Request

## Related issue
Fixes https://github.com/meilisearch/meilisearch/issues/4607

## What does this PR do?
- Ensure that anything falling under `_vectors` is NOT searchable, filterable or sortable
- [x] per embedder, add a roaring bitmap of documents that provide "userProvided" embeddings
- [x] in the indexing process in extract_vector_points, set the bit corresponding to the document depending on the "userProvided" subfield in the _vectors field.
- [x] in the document DB in typed chunks, when writing the _vectors field, remove all keys corresponding to an embedder

Co-authored-by: Tamo <tamo@meilisearch.com>
Co-authored-by: Louis Dureuil <louis@meilisearch.com>

Cargo.lock

@@ -2455,6 +2455,7 @@ name = "index-scheduler"
 version = "1.9.0"
 dependencies = [
  "anyhow",
+ "arroy",
  "big_s",
  "bincode",
  "crossbeam",
@@ -2465,6 +2466,7 @@ dependencies = [
  "file-store",
  "flate2",
  "insta",
+ "maplit",
  "meili-snap",
  "meilisearch-auth",
  "meilisearch-types",
@@ -5301,9 +5303,9 @@ dependencies = [
 
 [[package]]
 name = "tracing-actix-web"
-version = "0.7.10"
+version = "0.7.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fa069bd1503dd526ee793bb3fce408895136c95fc86d2edb2acf1c646d7f0684"
+checksum = "4ee9e39a66d9b615644893ffc1704d2a89b5b315b7fd0228ad3182ca9a306b19"
 dependencies = [
  "actix-web",
  "mutually_exclusive_features",

dump/src/reader/snapshots/dump__reader__test__import_dump_v6_with_vectors-6.snap

@@ -780,7 +780,7 @@ expression: document
           1.3484878540039063
         ]
       ],
-      "userProvided": false
+      "regenerate": true
     }
   }
 }

dump/src/reader/snapshots/dump__reader__test__import_dump_v6_with_vectors-7.snap

@@ -779,7 +779,7 @@ expression: document
           1.04031240940094
         ]
       ],
-      "userProvided": false
+      "regenerate": true
     }
   }
 }

dump/src/reader/v3/settings.rs

@@ -152,6 +152,7 @@ impl Settings<Unchecked> {
 }
 
 #[derive(Debug, Clone, Deserialize)]
+#[allow(dead_code)] // otherwise rustc complains that the fields go unused
 #[cfg_attr(test, derive(serde::Serialize))]
 #[serde(deny_unknown_fields)]
 #[serde(rename_all = "camelCase")]

dump/src/reader/v4/settings.rs

@@ -182,6 +182,7 @@ impl Settings<Unchecked> {
     }
 }
 
+#[allow(dead_code)] // otherwise rustc complains that the fields go unused
 #[derive(Debug, Clone, Deserialize)]
 #[cfg_attr(test, derive(serde::Serialize))]
 #[serde(deny_unknown_fields)]

dump/src/reader/v5/tasks.rs

@@ -200,6 +200,7 @@ impl std::ops::Deref for IndexUid {
     }
 }
 
+#[allow(dead_code)] // otherwise rustc complains that the fields go unused
 #[derive(Debug)]
 #[cfg_attr(test, derive(serde::Serialize))]
 #[cfg_attr(test, serde(rename_all = "camelCase"))]

index-scheduler/Cargo.toml

@@ -40,7 +40,9 @@ ureq = "2.9.7"
 uuid = { version = "1.6.1", features = ["serde", "v4"] }
 
 [dev-dependencies]
+arroy = "0.3.1"
 big_s = "1.0.2"
 crossbeam = "0.8.4"
 insta = { version = "1.34.0", features = ["json", "redactions"] }
+maplit = "1.0.2"
 meili-snap = { path = "../meili-snap" }

index-scheduler/src/batch.rs

@@ -909,6 +909,7 @@ impl IndexScheduler {
 
                     let fields_ids_map = index.fields_ids_map(&rtxn)?;
                     let all_fields: Vec<_> = fields_ids_map.iter().map(|(id, _)| id).collect();
+                    let embedding_configs = index.embedding_configs(&rtxn)?;
 
                     // 3.1. Dump the documents
                     for ret in index.all_documents(&rtxn)? {
@@ -951,16 +952,21 @@ impl IndexScheduler {
                             };
 
                             for (embedder_name, embeddings) in embeddings {
-                                // don't change the entry if it already exists, because it was user-provided
-                                vectors.entry(embedder_name).or_insert_with(|| {
-                                    let embeddings = ExplicitVectors {
-                                        embeddings: VectorOrArrayOfVectors::from_array_of_vectors(
-                                            embeddings,
-                                        ),
-                                        user_provided: false,
-                                    };
-                                    serde_json::to_value(embeddings).unwrap()
-                                });
+                                let user_provided = embedding_configs
+                                    .iter()
+                                    .find(|conf| conf.name == embedder_name)
+                                    .is_some_and(|conf| conf.user_provided.contains(id));
+
+                                let embeddings = ExplicitVectors {
+                                    embeddings: Some(
+                                        VectorOrArrayOfVectors::from_array_of_vectors(embeddings),
+                                    ),
+                                    regenerate: !user_provided,
+                                };
+                                vectors.insert(
+                                    embedder_name,
+                                    serde_json::to_value(embeddings).unwrap(),
+                                );
                             }
                         }
 

index-scheduler/src/lib.rs

@@ -53,6 +53,7 @@ use meilisearch_types::heed::byteorder::BE;
 use meilisearch_types::heed::types::{SerdeBincode, SerdeJson, Str, I128};
 use meilisearch_types::heed::{self, Database, Env, PutFlags, RoTxn, RwTxn};
 use meilisearch_types::milli::documents::DocumentsBatchBuilder;
+use meilisearch_types::milli::index::IndexEmbeddingConfig;
 use meilisearch_types::milli::update::IndexerConfig;
 use meilisearch_types::milli::vector::{Embedder, EmbedderOptions, EmbeddingConfigs};
 use meilisearch_types::milli::{self, CboRoaringBitmapCodec, Index, RoaringBitmapCodec, BEU32};
@@ -1459,33 +1460,39 @@ impl IndexScheduler {
     // TODO: consider using a type alias or a struct embedder/template
     pub fn embedders(
         &self,
-        embedding_configs: Vec<(String, milli::vector::EmbeddingConfig)>,
+        embedding_configs: Vec<IndexEmbeddingConfig>,
     ) -> Result<EmbeddingConfigs> {
         let res: Result<_> = embedding_configs
             .into_iter()
-            .map(|(name, milli::vector::EmbeddingConfig { embedder_options, prompt })| {
-                let prompt =
-                    Arc::new(prompt.try_into().map_err(meilisearch_types::milli::Error::from)?);
-                // optimistically return existing embedder
-                {
-                    let embedders = self.embedders.read().unwrap();
-                    if let Some(embedder) = embedders.get(&embedder_options) {
-                        return Ok((name, (embedder.clone(), prompt)));
+            .map(
+                |IndexEmbeddingConfig {
+                     name,
+                     config: milli::vector::EmbeddingConfig { embedder_options, prompt },
+                     ..
+                 }| {
+                    let prompt =
+                        Arc::new(prompt.try_into().map_err(meilisearch_types::milli::Error::from)?);
+                    // optimistically return existing embedder
+                    {
+                        let embedders = self.embedders.read().unwrap();
+                        if let Some(embedder) = embedders.get(&embedder_options) {
+                            return Ok((name, (embedder.clone(), prompt)));
+                        }
                     }
-                }
 
-                // add missing embedder
-                let embedder = Arc::new(
-                    Embedder::new(embedder_options.clone())
-                        .map_err(meilisearch_types::milli::vector::Error::from)
-                        .map_err(meilisearch_types::milli::Error::from)?,
-                );
-                {
-                    let mut embedders = self.embedders.write().unwrap();
-                    embedders.insert(embedder_options, embedder.clone());
-                }
-                Ok((name, (embedder, prompt)))
-            })
+                    // add missing embedder
+                    let embedder = Arc::new(
+                        Embedder::new(embedder_options.clone())
+                            .map_err(meilisearch_types::milli::vector::Error::from)
+                            .map_err(meilisearch_types::milli::Error::from)?,
+                    );
+                    {
+                        let mut embedders = self.embedders.write().unwrap();
+                        embedders.insert(embedder_options, embedder.clone());
+                    }
+                    Ok((name, (embedder, prompt)))
+                },
+            )
             .collect();
         res.map(EmbeddingConfigs::new)
     }
@@ -1748,6 +1755,9 @@ mod tests {
     use meilisearch_types::milli::update::IndexDocumentsMethod::{
         ReplaceDocuments, UpdateDocuments,
     };
+    use meilisearch_types::milli::update::Setting;
+    use meilisearch_types::milli::vector::settings::EmbeddingSettings;
+    use meilisearch_types::settings::Unchecked;
     use meilisearch_types::tasks::IndexSwap;
     use meilisearch_types::VERSION_FILE_NAME;
     use tempfile::{NamedTempFile, TempDir};
@@ -1826,6 +1836,7 @@ mod tests {
             assert_eq!(breakpoint, (Init, false));
             let index_scheduler_handle = IndexSchedulerHandle {
                 _tempdir: tempdir,
+                index_scheduler: index_scheduler.private_clone(),
                 test_breakpoint_rcv: receiver,
                 last_breakpoint: breakpoint.0,
             };
@@ -1914,6 +1925,7 @@ mod tests {
 
     pub struct IndexSchedulerHandle {
         _tempdir: TempDir,
+        index_scheduler: IndexScheduler,
         test_breakpoint_rcv: crossbeam::channel::Receiver<(Breakpoint, bool)>,
         last_breakpoint: Breakpoint,
     }
@@ -1931,9 +1943,13 @@ mod tests {
             {
                 Ok(b) => b,
                 Err(RecvTimeoutError::Timeout) => {
-                    panic!("The scheduler seems to be waiting for a new task while your test is waiting for a breakpoint.")
+                    let state = snapshot_index_scheduler(&self.index_scheduler);
+                    panic!("The scheduler seems to be waiting for a new task while your test is waiting for a breakpoint.\n{state}")
+                }
+                Err(RecvTimeoutError::Disconnected) => {
+                    let state = snapshot_index_scheduler(&self.index_scheduler);
+                    panic!("The scheduler crashed.\n{state}")
                 }
-                Err(RecvTimeoutError::Disconnected) => panic!("The scheduler crashed."),
             };
             // if we've already encountered a breakpoint we're supposed to be stuck on the false
             // and we expect the same variant with the true to come now.
@@ -1952,9 +1968,13 @@ mod tests {
             {
                 Ok(b) => b,
                 Err(RecvTimeoutError::Timeout) => {
-                    panic!("The scheduler seems to be waiting for a new task while your test is waiting for a breakpoint.")
+                    let state = snapshot_index_scheduler(&self.index_scheduler);
+                    panic!("The scheduler seems to be waiting for a new task while your test is waiting for a breakpoint.\n{state}")
+                }
+                Err(RecvTimeoutError::Disconnected) => {
+                    let state = snapshot_index_scheduler(&self.index_scheduler);
+                    panic!("The scheduler crashed.\n{state}")
                 }
-                Err(RecvTimeoutError::Disconnected) => panic!("The scheduler crashed."),
             };
             assert!(!b, "Found the breakpoint handle in a bad state. Check your test suite");
 
@@ -1968,9 +1988,10 @@ mod tests {
         fn advance_till(&mut self, breakpoints: impl IntoIterator<Item = Breakpoint>) {
             for breakpoint in breakpoints {
                 let b = self.advance();
+                let state = snapshot_index_scheduler(&self.index_scheduler);
                 assert_eq!(
                     b, breakpoint,
-                    "Was expecting the breakpoint `{:?}` but instead got `{:?}`.",
+                    "Was expecting the breakpoint `{:?}` but instead got `{:?}`.\n{state}",
                     breakpoint, b
                 );
             }
@@ -1995,6 +2016,7 @@ mod tests {
         // Wait for one successful batch.
         #[track_caller]
         fn advance_one_successful_batch(&mut self) {
+            self.index_scheduler.assert_internally_consistent();
             self.advance_till([Start, BatchCreated]);
             loop {
                 match self.advance() {
@@ -2003,13 +2025,17 @@ mod tests {
                     InsideProcessBatch => (),
                     // the batch went successfully, we can stop the loop and go on with the next states.
                     ProcessBatchSucceeded => break,
-                    AbortedIndexation => panic!("The batch was aborted."),
-                    ProcessBatchFailed => panic!("The batch failed."),
+                    AbortedIndexation => panic!("The batch was aborted.\n{}", snapshot_index_scheduler(&self.index_scheduler)),
+                    ProcessBatchFailed => {
+                        while self.advance() != Start {}
+                        panic!("The batch failed.\n{}", snapshot_index_scheduler(&self.index_scheduler))
+                    },
                     breakpoint => panic!("Encountered an impossible breakpoint `{:?}`, this is probably an issue with the test suite.", breakpoint),
                 }
             }
 
             self.advance_till([AfterProcessing]);
+            self.index_scheduler.assert_internally_consistent();
         }
 
         // Wait for one failed batch.
@@ -2023,8 +2049,8 @@ mod tests {
                     InsideProcessBatch => (),
                     // the batch went failed, we can stop the loop and go on with the next states.
                     ProcessBatchFailed => break,
-                    ProcessBatchSucceeded => panic!("The batch succeeded. (and it wasn't supposed to sorry)"),
-                    AbortedIndexation => panic!("The batch was aborted."),
+                    ProcessBatchSucceeded => panic!("The batch succeeded. (and it wasn't supposed to sorry)\n{}", snapshot_index_scheduler(&self.index_scheduler)),
+                    AbortedIndexation => panic!("The batch was aborted.\n{}", snapshot_index_scheduler(&self.index_scheduler)),
                     breakpoint => panic!("Encountered an impossible breakpoint `{:?}`, this is probably an issue with the test suite.", breakpoint),
                 }
             }
@@ -3052,8 +3078,10 @@ mod tests {
         let rtxn = index.read_txn().unwrap();
 
         let configs = index.embedding_configs(&rtxn).unwrap();
-        let (_, embedding_config) = configs.first().unwrap();
-        insta::assert_json_snapshot!(embedding_config.embedder_options);
+        let IndexEmbeddingConfig { name, config, user_provided } = configs.first().unwrap();
+        insta::assert_snapshot!(name, @"default");
+        insta::assert_debug_snapshot!(user_provided, @"RoaringBitmap<[]>");
+        insta::assert_json_snapshot!(config.embedder_options);
     }
 
     #[test]
@@ -4989,7 +5017,6 @@ mod tests {
                 false,
             )
             .unwrap();
-        index_scheduler.assert_internally_consistent();
 
         snapshot!(snapshot_index_scheduler(&index_scheduler), name: "after_registering_settings_task_vectors");
 
@@ -5000,7 +5027,7 @@ mod tests {
             insta::assert_json_snapshot!(task.details);
         }
 
-        handle.advance_n_successful_batches(1);
+        handle.advance_one_successful_batch();
         snapshot!(snapshot_index_scheduler(&index_scheduler), name: "settings_update_processed_vectors");
 
         {
@@ -5017,13 +5044,17 @@ mod tests {
             let configs = index.embedding_configs(&rtxn).unwrap();
             // for consistency with the below
             #[allow(clippy::get_first)]
-            let (name, fakerest_config) = configs.get(0).unwrap();
-            insta::assert_json_snapshot!(name, @r###""A_fakerest""###);
+            let IndexEmbeddingConfig { name, config: fakerest_config, user_provided } =
+                configs.get(0).unwrap();
+            insta::assert_snapshot!(name, @"A_fakerest");
+            insta::assert_debug_snapshot!(user_provided, @"RoaringBitmap<[]>");
             insta::assert_json_snapshot!(fakerest_config.embedder_options);
             let fakerest_name = name.clone();
 
-            let (name, simple_hf_config) = configs.get(1).unwrap();
-            insta::assert_json_snapshot!(name, @r###""B_small_hf""###);
+            let IndexEmbeddingConfig { name, config: simple_hf_config, user_provided } =
+                configs.get(1).unwrap();
+            insta::assert_snapshot!(name, @"B_small_hf");
+            insta::assert_debug_snapshot!(user_provided, @"RoaringBitmap<[]>");
             insta::assert_json_snapshot!(simple_hf_config.embedder_options);
             let simple_hf_name = name.clone();
 
@@ -5038,25 +5069,25 @@ mod tests {
         // add one doc, specifying vectors
 
         let doc = serde_json::json!(
-                    {
-                        "id": 0,
-                        "doggo": "Intel",
-                        "breed": "beagle",
-                        "_vectors": {
-                            &fakerest_name: {
-                                // this will never trigger regeneration, which is good because we can't actually generate with
-                                // this embedder
-                                "userProvided": true,
-                                "embeddings": beagle_embed,
-                            },
-                            &simple_hf_name: {
-                                // this will be regenerated on updates
-                                "userProvided": false,
-                                "embeddings": lab_embed,
-                            },
-                            "noise": [0.1, 0.2, 0.3]
-                        }
-                    }
+            {
+                "id": 0,
+                "doggo": "Intel",
+                "breed": "beagle",
+                "_vectors": {
+                    &fakerest_name: {
+                        // this will never trigger regeneration, which is good because we can't actually generate with
+                        // this embedder
+                        "regenerate": false,
+                        "embeddings": beagle_embed,
+                    },
+                    &simple_hf_name: {
+                        // this will be regenerated on updates
+                        "regenerate": true,
+                        "embeddings": lab_embed,
+                    },
+                    "noise": [0.1, 0.2, 0.3]
+                }
+            }
         );
 
         let (uuid, mut file) = index_scheduler.create_update_file_with_uuid(0u128).unwrap();
@@ -5078,7 +5109,6 @@ mod tests {
                 false,
             )
             .unwrap();
-        index_scheduler.assert_internally_consistent();
 
         snapshot!(snapshot_index_scheduler(&index_scheduler), name: "after adding Intel");
 
@@ -5091,6 +5121,19 @@ mod tests {
             let index = index_scheduler.index("doggos").unwrap();
             let rtxn = index.read_txn().unwrap();
 
+            // Ensure the document have been inserted into the relevant bitamp
+            let configs = index.embedding_configs(&rtxn).unwrap();
+            // for consistency with the below
+            #[allow(clippy::get_first)]
+            let IndexEmbeddingConfig { name, config: _, user_provided: user_defined } =
+                configs.get(0).unwrap();
+            insta::assert_snapshot!(name, @"A_fakerest");
+            insta::assert_debug_snapshot!(user_defined, @"RoaringBitmap<[0]>");
+
+            let IndexEmbeddingConfig { name, config: _, user_provided } = configs.get(1).unwrap();
+            insta::assert_snapshot!(name, @"B_small_hf");
+            insta::assert_debug_snapshot!(user_provided, @"RoaringBitmap<[]>");
+
             let embeddings = index.embeddings(&rtxn, 0).unwrap();
 
             assert_json_snapshot!(embeddings[&simple_hf_name][0] == lab_embed, @"true");
@@ -5140,7 +5183,6 @@ mod tests {
                 false,
             )
             .unwrap();
-        index_scheduler.assert_internally_consistent();
 
         snapshot!(snapshot_index_scheduler(&index_scheduler), name: "Intel to kefir");
 
@@ -5153,11 +5195,25 @@ mod tests {
                 let index = index_scheduler.index("doggos").unwrap();
                 let rtxn = index.read_txn().unwrap();
 
+                // Ensure the document have been inserted into the relevant bitamp
+                let configs = index.embedding_configs(&rtxn).unwrap();
+                // for consistency with the below
+                #[allow(clippy::get_first)]
+                let IndexEmbeddingConfig { name, config: _, user_provided: user_defined } =
+                    configs.get(0).unwrap();
+                insta::assert_snapshot!(name, @"A_fakerest");
+                insta::assert_debug_snapshot!(user_defined, @"RoaringBitmap<[0]>");
+
+                let IndexEmbeddingConfig { name, config: _, user_provided } =
+                    configs.get(1).unwrap();
+                insta::assert_snapshot!(name, @"B_small_hf");
+                insta::assert_debug_snapshot!(user_provided, @"RoaringBitmap<[]>");
+
                 let embeddings = index.embeddings(&rtxn, 0).unwrap();
 
-                // automatically changed to patou
+                // automatically changed to patou because set to regenerate
                 assert_json_snapshot!(embeddings[&simple_hf_name][0] == patou_embed, @"true");
-                // remained beagle because set to userProvided
+                // remained beagle
                 assert_json_snapshot!(embeddings[&fakerest_name][0] == beagle_embed, @"true");
 
                 let doc = index.documents(&rtxn, std::iter::once(0)).unwrap()[0].1;
@@ -5176,4 +5232,578 @@ mod tests {
             }
         }
     }
+
+    #[test]
+    fn import_vectors_first_and_embedder_later() {
+        let (index_scheduler, mut handle) = IndexScheduler::test(true, vec![]);
+
+        let content = serde_json::json!(
+            [
+                {
+                    "id": 0,
+                    "doggo": "kefir",
+                },
+                {
+                    "id": 1,
+                    "doggo": "intel",
+                    "_vectors": {
+                        "my_doggo_embedder": vec![1; 384],
+                        "unknown embedder": vec![1, 2, 3],
+                    }
+                },
+                {
+                    "id": 2,
+                    "doggo": "max",
+                    "_vectors": {
+                        "my_doggo_embedder": {
+                            "regenerate": false,
+                            "embeddings": vec![2; 384],
+                        },
+                        "unknown embedder": vec![4, 5],
+                    },
+                },
+                {
+                    "id": 3,
+                    "doggo": "marcel",
+                    "_vectors": {
+                        "my_doggo_embedder": {
+                            "regenerate": true,
+                            "embeddings": vec![3; 384],
+                        },
+                    },
+                },
+                {
+                    "id": 4,
+                    "doggo": "sora",
+                    "_vectors": {
+                        "my_doggo_embedder": {
+                            "regenerate": true,
+                        },
+                    },
+                },
+            ]
+        );
+
+        let (uuid, mut file) = index_scheduler.create_update_file_with_uuid(0_u128).unwrap();
+        let documents_count =
+            read_json(serde_json::to_string_pretty(&content).unwrap().as_bytes(), &mut file)
+                .unwrap();
+        snapshot!(documents_count, @"5");
+        file.persist().unwrap();
+
+        index_scheduler
+            .register(
+                KindWithContent::DocumentAdditionOrUpdate {
+                    index_uid: S("doggos"),
+                    primary_key: None,
+                    method: ReplaceDocuments,
+                    content_file: uuid,
+                    documents_count,
+                    allow_index_creation: true,
+                },
+                None,
+                false,
+            )
+            .unwrap();
+        handle.advance_one_successful_batch();
+
+        let index = index_scheduler.index("doggos").unwrap();
+        let rtxn = index.read_txn().unwrap();
+        let field_ids_map = index.fields_ids_map(&rtxn).unwrap();
+        let field_ids = field_ids_map.ids().collect::<Vec<_>>();
+        let documents = index
+            .all_documents(&rtxn)
+            .unwrap()
+            .map(|ret| obkv_to_json(&field_ids, &field_ids_map, ret.unwrap().1).unwrap())
+            .collect::<Vec<_>>();
+        snapshot!(serde_json::to_string(&documents).unwrap(), name: "documents after initial push");
+
+        let setting = meilisearch_types::settings::Settings::<Unchecked> {
+            embedders: Setting::Set(maplit::btreemap! {
+                S("my_doggo_embedder") => Setting::Set(EmbeddingSettings {
+                    source: Setting::Set(milli::vector::settings::EmbedderSource::HuggingFace),
+                    model: Setting::Set(S("sentence-transformers/all-MiniLM-L6-v2")),
+                    revision: Setting::Set(S("e4ce9877abf3edfe10b0d82785e83bdcb973e22e")),
+                    document_template: Setting::Set(S("{{doc.doggo}}")),
+                    ..Default::default()
+                })
+            }),
+            ..Default::default()
+        };
+        index_scheduler
+            .register(
+                KindWithContent::SettingsUpdate {
+                    index_uid: S("doggos"),
+                    new_settings: Box::new(setting),
+                    is_deletion: false,
+                    allow_index_creation: false,
+                },
+                None,
+                false,
+            )
+            .unwrap();
+        index_scheduler.assert_internally_consistent();
+        handle.advance_one_successful_batch();
+        index_scheduler.assert_internally_consistent();
+
+        let index = index_scheduler.index("doggos").unwrap();
+        let rtxn = index.read_txn().unwrap();
+        let field_ids_map = index.fields_ids_map(&rtxn).unwrap();
+        let field_ids = field_ids_map.ids().collect::<Vec<_>>();
+        let documents = index
+            .all_documents(&rtxn)
+            .unwrap()
+            .map(|ret| obkv_to_json(&field_ids, &field_ids_map, ret.unwrap().1).unwrap())
+            .collect::<Vec<_>>();
+        // the all the vectors linked to the new specified embedder have been removed
+        // Only the unknown embedders stays in the document DB
+        snapshot!(serde_json::to_string(&documents).unwrap(), @r###"[{"id":0,"doggo":"kefir"},{"id":1,"doggo":"intel","_vectors":{"unknown embedder":[1.0,2.0,3.0]}},{"id":2,"doggo":"max","_vectors":{"unknown embedder":[4.0,5.0]}},{"id":3,"doggo":"marcel"},{"id":4,"doggo":"sora"}]"###);
+        let conf = index.embedding_configs(&rtxn).unwrap();
+        // even though we specified the vector for the ID 3, it shouldn't be marked
+        // as user provided since we explicitely marked it as NOT user provided.
+        snapshot!(format!("{conf:#?}"), @r###"
+        [
+            IndexEmbeddingConfig {
+                name: "my_doggo_embedder",
+                config: EmbeddingConfig {
+                    embedder_options: HuggingFace(
+                        EmbedderOptions {
+                            model: "sentence-transformers/all-MiniLM-L6-v2",
+                            revision: Some(
+                                "e4ce9877abf3edfe10b0d82785e83bdcb973e22e",
+                            ),
+                            distribution: None,
+                        },
+                    ),
+                    prompt: PromptData {
+                        template: "{{doc.doggo}}",
+                    },
+                },
+                user_provided: RoaringBitmap<[1, 2]>,
+            },
+        ]
+        "###);
+        let docid = index.external_documents_ids.get(&rtxn, "0").unwrap().unwrap();
+        let embeddings = index.embeddings(&rtxn, docid).unwrap();
+        let embedding = &embeddings["my_doggo_embedder"];
+        assert!(!embedding.is_empty(), "{embedding:?}");
+
+        // the document with the id 3 should keep its original embedding
+        let docid = index.external_documents_ids.get(&rtxn, "3").unwrap().unwrap();
+        let mut embeddings = Vec::new();
+
+        'vectors: for i in 0..=u8::MAX {
+            let reader = arroy::Reader::open(&rtxn, i as u16, index.vector_arroy)
+                .map(Some)
+                .or_else(|e| match e {
+                    arroy::Error::MissingMetadata => Ok(None),
+                    e => Err(e),
+                })
+                .transpose();
+
+            let Some(reader) = reader else {
+                break 'vectors;
+            };
+
+            let embedding = reader.unwrap().item_vector(&rtxn, docid).unwrap();
+            if let Some(embedding) = embedding {
+                embeddings.push(embedding)
+            } else {
+                break 'vectors;
+            }
+        }
+
+        snapshot!(embeddings.len(), @"1");
+        assert!(embeddings[0].iter().all(|i| *i == 3.0), "{:?}", embeddings[0]);
+
+        // If we update marcel it should regenerate its embedding automatically
+
+        let content = serde_json::json!(
+            [
+                {
+                    "id": 3,
+                    "doggo": "marvel",
+                },
+                {
+                    "id": 4,
+                    "doggo": "sorry",
+                },
+            ]
+        );
+
+        let (uuid, mut file) = index_scheduler.create_update_file_with_uuid(1_u128).unwrap();
+        let documents_count =
+            read_json(serde_json::to_string_pretty(&content).unwrap().as_bytes(), &mut file)
+                .unwrap();
+        snapshot!(documents_count, @"2");
+        file.persist().unwrap();
+
+        index_scheduler
+            .register(
+                KindWithContent::DocumentAdditionOrUpdate {
+                    index_uid: S("doggos"),
+                    primary_key: None,
+                    method: UpdateDocuments,
+                    content_file: uuid,
+                    documents_count,
+                    allow_index_creation: true,
+                },
+                None,
+                false,
+            )
+            .unwrap();
+        handle.advance_one_successful_batch();
+
+        // the document with the id 3 should have its original embedding updated
+        let rtxn = index.read_txn().unwrap();
+        let docid = index.external_documents_ids.get(&rtxn, "3").unwrap().unwrap();
+        let doc = index.documents(&rtxn, Some(docid)).unwrap()[0];
+        let doc = obkv_to_json(&field_ids, &field_ids_map, doc.1).unwrap();
+        snapshot!(json_string!(doc), @r###"
+        {
+          "id": 3,
+          "doggo": "marvel"
+        }
+        "###);
+
+        let embeddings = index.embeddings(&rtxn, docid).unwrap();
+        let embedding = &embeddings["my_doggo_embedder"];
+
+        assert!(!embedding.is_empty());
+        assert!(!embedding[0].iter().all(|i| *i == 3.0), "{:?}", embedding[0]);
+
+        // the document with the id 4 should generate an embedding
+        let docid = index.external_documents_ids.get(&rtxn, "4").unwrap().unwrap();
+        let embeddings = index.embeddings(&rtxn, docid).unwrap();
+        let embedding = &embeddings["my_doggo_embedder"];
+
+        assert!(!embedding.is_empty());
+    }
+
+    #[test]
+    fn delete_document_containing_vector() {
+        // 1. Add an embedder
+        // 2. Push two documents containing a simple vector
+        // 3. Delete the first document
+        // 4. The user defined roaring bitmap shouldn't contains the id of the first document anymore
+        // 5. Clear the index
+        // 6. The user defined roaring bitmap shouldn't contains the id of the second document
+        let (index_scheduler, mut handle) = IndexScheduler::test(true, vec![]);
+
+        let setting = meilisearch_types::settings::Settings::<Unchecked> {
+            embedders: Setting::Set(maplit::btreemap! {
+                S("manual") => Setting::Set(EmbeddingSettings {
+                    source: Setting::Set(milli::vector::settings::EmbedderSource::UserProvided),
+                    dimensions: Setting::Set(3),
+                    ..Default::default()
+                })
+            }),
+            ..Default::default()
+        };
+        index_scheduler
+            .register(
+                KindWithContent::SettingsUpdate {
+                    index_uid: S("doggos"),
+                    new_settings: Box::new(setting),
+                    is_deletion: false,
+                    allow_index_creation: true,
+                },
+                None,
+                false,
+            )
+            .unwrap();
+        handle.advance_one_successful_batch();
+
+        let content = serde_json::json!(
+            [
+                {
+                    "id": 0,
+                    "doggo": "kefir",
+                    "_vectors": {
+                        "manual": vec![0, 0, 0],
+                    }
+                },
+                {
+                    "id": 1,
+                    "doggo": "intel",
+                    "_vectors": {
+                        "manual": vec![1, 1, 1],
+                    }
+                },
+            ]
+        );
+
+        let (uuid, mut file) = index_scheduler.create_update_file_with_uuid(0_u128).unwrap();
+        let documents_count =
+            read_json(serde_json::to_string_pretty(&content).unwrap().as_bytes(), &mut file)
+                .unwrap();
+        snapshot!(documents_count, @"2");
+        file.persist().unwrap();
+
+        index_scheduler
+            .register(
+                KindWithContent::DocumentAdditionOrUpdate {
+                    index_uid: S("doggos"),
+                    primary_key: None,
+                    method: ReplaceDocuments,
+                    content_file: uuid,
+                    documents_count,
+                    allow_index_creation: false,
+                },
+                None,
+                false,
+            )
+            .unwrap();
+        handle.advance_one_successful_batch();
+
+        index_scheduler
+            .register(
+                KindWithContent::DocumentDeletion {
+                    index_uid: S("doggos"),
+                    documents_ids: vec![S("1")],
+                },
+                None,
+                false,
+            )
+            .unwrap();
+        handle.advance_one_successful_batch();
+
+        let index = index_scheduler.index("doggos").unwrap();
+        let rtxn = index.read_txn().unwrap();
+        let field_ids_map = index.fields_ids_map(&rtxn).unwrap();
+        let field_ids = field_ids_map.ids().collect::<Vec<_>>();
+        let documents = index
+            .all_documents(&rtxn)
+            .unwrap()
+            .map(|ret| obkv_to_json(&field_ids, &field_ids_map, ret.unwrap().1).unwrap())
+            .collect::<Vec<_>>();
+        snapshot!(serde_json::to_string(&documents).unwrap(), @r###"[{"id":0,"doggo":"kefir"}]"###);
+        let conf = index.embedding_configs(&rtxn).unwrap();
+        snapshot!(format!("{conf:#?}"), @r###"
+        [
+            IndexEmbeddingConfig {
+                name: "manual",
+                config: EmbeddingConfig {
+                    embedder_options: UserProvided(
+                        EmbedderOptions {
+                            dimensions: 3,
+                            distribution: None,
+                        },
+                    ),
+                    prompt: PromptData {
+                        template: "{% for field in fields %} {{ field.name }}: {{ field.value }}\n{% endfor %}",
+                    },
+                },
+                user_provided: RoaringBitmap<[0]>,
+            },
+        ]
+        "###);
+        let docid = index.external_documents_ids.get(&rtxn, "0").unwrap().unwrap();
+        let embeddings = index.embeddings(&rtxn, docid).unwrap();
+        let embedding = &embeddings["manual"];
+        assert!(!embedding.is_empty(), "{embedding:?}");
+
+        index_scheduler
+            .register(KindWithContent::DocumentClear { index_uid: S("doggos") }, None, false)
+            .unwrap();
+        handle.advance_one_successful_batch();
+
+        let index = index_scheduler.index("doggos").unwrap();
+        let rtxn = index.read_txn().unwrap();
+        let field_ids_map = index.fields_ids_map(&rtxn).unwrap();
+        let field_ids = field_ids_map.ids().collect::<Vec<_>>();
+        let documents = index
+            .all_documents(&rtxn)
+            .unwrap()
+            .map(|ret| obkv_to_json(&field_ids, &field_ids_map, ret.unwrap().1).unwrap())
+            .collect::<Vec<_>>();
+        snapshot!(serde_json::to_string(&documents).unwrap(), @"[]");
+        let conf = index.embedding_configs(&rtxn).unwrap();
+        snapshot!(format!("{conf:#?}"), @r###"
+        [
+            IndexEmbeddingConfig {
+                name: "manual",
+                config: EmbeddingConfig {
+                    embedder_options: UserProvided(
+                        EmbedderOptions {
+                            dimensions: 3,
+                            distribution: None,
+                        },
+                    ),
+                    prompt: PromptData {
+                        template: "{% for field in fields %} {{ field.name }}: {{ field.value }}\n{% endfor %}",
+                    },
+                },
+                user_provided: RoaringBitmap<[]>,
+            },
+        ]
+        "###);
+    }
+
+    #[test]
+    fn delete_embedder_with_user_provided_vectors() {
+        // 1. Add two embedders
+        // 2. Push two documents containing a simple vector
+        // 3. The documents must not contain the vectors after the update as they are in the vectors db
+        // 3. Delete the embedders
+        // 4. The documents contain the vectors again
+        let (index_scheduler, mut handle) = IndexScheduler::test(true, vec![]);
+
+        let setting = meilisearch_types::settings::Settings::<Unchecked> {
+            embedders: Setting::Set(maplit::btreemap! {
+                S("manual") => Setting::Set(EmbeddingSettings {
+                    source: Setting::Set(milli::vector::settings::EmbedderSource::UserProvided),
+                    dimensions: Setting::Set(3),
+                    ..Default::default()
+                }),
+                S("my_doggo_embedder") => Setting::Set(EmbeddingSettings {
+                    source: Setting::Set(milli::vector::settings::EmbedderSource::HuggingFace),
+                    model: Setting::Set(S("sentence-transformers/all-MiniLM-L6-v2")),
+                    revision: Setting::Set(S("e4ce9877abf3edfe10b0d82785e83bdcb973e22e")),
+                    document_template: Setting::Set(S("{{doc.doggo}}")),
+                    ..Default::default()
+                }),
+            }),
+            ..Default::default()
+        };
+        index_scheduler
+            .register(
+                KindWithContent::SettingsUpdate {
+                    index_uid: S("doggos"),
+                    new_settings: Box::new(setting),
+                    is_deletion: false,
+                    allow_index_creation: true,
+                },
+                None,
+                false,
+            )
+            .unwrap();
+        handle.advance_one_successful_batch();
+
+        let content = serde_json::json!(
+            [
+                {
+                    "id": 0,
+                    "doggo": "kefir",
+                    "_vectors": {
+                        "manual": vec![0, 0, 0],
+                        "my_doggo_embedder": vec![1; 384],
+                    }
+                },
+                {
+                    "id": 1,
+                    "doggo": "intel",
+                    "_vectors": {
+                        "manual": vec![1, 1, 1],
+                    }
+                },
+            ]
+        );
+
+        let (uuid, mut file) = index_scheduler.create_update_file_with_uuid(0_u128).unwrap();
+        let documents_count =
+            read_json(serde_json::to_string_pretty(&content).unwrap().as_bytes(), &mut file)
+                .unwrap();
+        snapshot!(documents_count, @"2");
+        file.persist().unwrap();
+
+        index_scheduler
+            .register(
+                KindWithContent::DocumentAdditionOrUpdate {
+                    index_uid: S("doggos"),
+                    primary_key: None,
+                    method: ReplaceDocuments,
+                    content_file: uuid,
+                    documents_count,
+                    allow_index_creation: false,
+                },
+                None,
+                false,
+            )
+            .unwrap();
+        handle.advance_one_successful_batch();
+
+        {
+            let index = index_scheduler.index("doggos").unwrap();
+            let rtxn = index.read_txn().unwrap();
+            let field_ids_map = index.fields_ids_map(&rtxn).unwrap();
+            let field_ids = field_ids_map.ids().collect::<Vec<_>>();
+            let documents = index
+                .all_documents(&rtxn)
+                .unwrap()
+                .map(|ret| obkv_to_json(&field_ids, &field_ids_map, ret.unwrap().1).unwrap())
+                .collect::<Vec<_>>();
+            snapshot!(serde_json::to_string(&documents).unwrap(), @r###"[{"id":0,"doggo":"kefir"},{"id":1,"doggo":"intel"}]"###);
+        }
+
+        {
+            let setting = meilisearch_types::settings::Settings::<Unchecked> {
+                embedders: Setting::Set(maplit::btreemap! {
+                    S("manual") => Setting::Reset,
+                }),
+                ..Default::default()
+            };
+            index_scheduler
+                .register(
+                    KindWithContent::SettingsUpdate {
+                        index_uid: S("doggos"),
+                        new_settings: Box::new(setting),
+                        is_deletion: false,
+                        allow_index_creation: true,
+                    },
+                    None,
+                    false,
+                )
+                .unwrap();
+            handle.advance_one_successful_batch();
+        }
+
+        {
+            let index = index_scheduler.index("doggos").unwrap();
+            let rtxn = index.read_txn().unwrap();
+            let field_ids_map = index.fields_ids_map(&rtxn).unwrap();
+            let field_ids = field_ids_map.ids().collect::<Vec<_>>();
+            let documents = index
+                .all_documents(&rtxn)
+                .unwrap()
+                .map(|ret| obkv_to_json(&field_ids, &field_ids_map, ret.unwrap().1).unwrap())
+                .collect::<Vec<_>>();
+            snapshot!(serde_json::to_string(&documents).unwrap(), @r###"[{"id":0,"doggo":"kefir","_vectors":{"manual":{"embeddings":[[0.0,0.0,0.0]],"regenerate":false}}},{"id":1,"doggo":"intel","_vectors":{"manual":{"embeddings":[[1.0,1.0,1.0]],"regenerate":false}}}]"###);
+        }
+
+        {
+            let setting = meilisearch_types::settings::Settings::<Unchecked> {
+                embedders: Setting::Reset,
+                ..Default::default()
+            };
+            index_scheduler
+                .register(
+                    KindWithContent::SettingsUpdate {
+                        index_uid: S("doggos"),
+                        new_settings: Box::new(setting),
+                        is_deletion: false,
+                        allow_index_creation: true,
+                    },
+                    None,
+                    false,
+                )
+                .unwrap();
+            handle.advance_one_successful_batch();
+        }
+
+        {
+            let index = index_scheduler.index("doggos").unwrap();
+            let rtxn = index.read_txn().unwrap();
+            let field_ids_map = index.fields_ids_map(&rtxn).unwrap();
+            let field_ids = field_ids_map.ids().collect::<Vec<_>>();
+            let documents = index
+                .all_documents(&rtxn)
+                .unwrap()
+                .map(|ret| obkv_to_json(&field_ids, &field_ids_map, ret.unwrap().1).unwrap())
+                .collect::<Vec<_>>();
+
+            // FIXME: redaction
+            snapshot!(json_string!(serde_json::to_string(&documents).unwrap(), { "[]._vectors.doggo_embedder.embeddings" => "[vector]" }),  @r###""[{\"id\":0,\"doggo\":\"kefir\",\"_vectors\":{\"manual\":{\"embeddings\":[[0.0,0.0,0.0]],\"regenerate\":false},\"my_doggo_embedder\":{\"embeddings\":[[1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0]],\"regenerate\":false}}},{\"id\":1,\"doggo\":\"intel\",\"_vectors\":{\"manual\":{\"embeddings\":[[1.0,1.0,1.0]],\"regenerate\":false}}}]""###);
+        }
+    }
 }

index-scheduler/src/snapshots/index_scheduler__tests__import_vectors-15.snap

@@ -6,10 +6,6 @@ expression: doc
   "doggo": "Intel",
   "breed": "beagle",
   "_vectors": {
-    "A_fakerest": {
-      "embeddings": "[vector]",
-      "userProvided": true
-    },
     "noise": [
       0.1,
       0.2,

index-scheduler/src/snapshots/index_scheduler__tests__import_vectors-22.snap

@@ -6,10 +6,6 @@ expression: doc
   "doggo": "kefir",
   "breed": "patou",
   "_vectors": {
-    "A_fakerest": {
-      "embeddings": "[vector]",
-      "userProvided": true
-    },
     "noise": [
       0.1,
       0.2,

meilisearch-types/src/error.rs

@@ -222,6 +222,7 @@ InvalidApiKeyUid                      , InvalidRequest       , BAD_REQUEST ;
 InvalidContentType                    , InvalidRequest       , UNSUPPORTED_MEDIA_TYPE ;
 InvalidDocumentCsvDelimiter           , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentFields                 , InvalidRequest       , BAD_REQUEST ;
+InvalidDocumentRetrieveVectors        , InvalidRequest       , BAD_REQUEST ;
 MissingDocumentFilter                 , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentFilter                 , InvalidRequest       , BAD_REQUEST ;
 InvalidDocumentGeoField               , InvalidRequest       , BAD_REQUEST ;
@@ -240,9 +241,11 @@ InvalidSearchAttributesToSearchOn     , InvalidRequest       , BAD_REQUEST ;
 InvalidSearchAttributesToCrop         , InvalidRequest       , BAD_REQUEST ;
 InvalidSearchAttributesToHighlight    , InvalidRequest       , BAD_REQUEST ;
 InvalidSimilarAttributesToRetrieve    , InvalidRequest       , BAD_REQUEST ;
+InvalidSimilarRetrieveVectors         , InvalidRequest       , BAD_REQUEST ;
 InvalidSearchAttributesToRetrieve     , InvalidRequest       , BAD_REQUEST ;
 InvalidSearchRankingScoreThreshold    , InvalidRequest       , BAD_REQUEST ;
 InvalidSimilarRankingScoreThreshold   , InvalidRequest       , BAD_REQUEST ;
+InvalidSearchRetrieveVectors          , InvalidRequest       , BAD_REQUEST ;
 InvalidSearchCropLength               , InvalidRequest       , BAD_REQUEST ;
 InvalidSearchCropMarker               , InvalidRequest       , BAD_REQUEST ;
 InvalidSearchFacets                   , InvalidRequest       , BAD_REQUEST ;

meilisearch-types/src/settings.rs

@@ -8,6 +8,7 @@ use std::str::FromStr;
 
 use deserr::{DeserializeError, Deserr, ErrorKind, MergeWithError, ValuePointerRef};
 use fst::IntoStreamer;
+use milli::index::IndexEmbeddingConfig;
 use milli::proximity::ProximityPrecision;
 use milli::update::Setting;
 use milli::{Criterion, CriterionError, Index, DEFAULT_VALUES_PER_FACET};
@@ -672,7 +673,7 @@ pub fn settings(
     let embedders: BTreeMap<_, _> = index
         .embedding_configs(rtxn)?
         .into_iter()
-        .map(|(name, config)| (name, Setting::Set(config.into())))
+        .map(|IndexEmbeddingConfig { name, config, .. }| (name, Setting::Set(config.into())))
         .collect();
     let embedders = if embedders.is_empty() { Setting::NotSet } else { Setting::Set(embedders) };
 

meilisearch/src/analytics/mod.rs

@@ -74,8 +74,8 @@ pub enum DocumentDeletionKind {
 
 #[derive(Copy, Clone, Debug, PartialEq, Eq)]
 pub enum DocumentFetchKind {
-    PerDocumentId,
-    Normal { with_filter: bool, limit: usize, offset: usize },
+    PerDocumentId { retrieve_vectors: bool },
+    Normal { with_filter: bool, limit: usize, offset: usize, retrieve_vectors: bool },
 }
 
 pub trait Analytics: Sync + Send {

meilisearch/src/analytics/segment_analytics.rs

@@ -622,6 +622,7 @@ pub struct SearchAggregator {
     // Whether a non-default embedder was specified
     embedder: bool,
     hybrid: bool,
+    retrieve_vectors: bool,
 
     // every time a search is done, we increment the counter linked to the used settings
     matching_strategy: HashMap<String, usize>,
@@ -662,6 +663,7 @@ impl SearchAggregator {
             page,
             hits_per_page,
             attributes_to_retrieve: _,
+            retrieve_vectors,
             attributes_to_crop: _,
             crop_length,
             attributes_to_highlight: _,
@@ -728,6 +730,7 @@ impl SearchAggregator {
         if let Some(ref vector) = vector {
             ret.max_vector_size = vector.len();
         }
+        ret.retrieve_vectors |= retrieve_vectors;
 
         if query.is_finite_pagination() {
             let limit = hits_per_page.unwrap_or_else(DEFAULT_SEARCH_LIMIT);
@@ -803,6 +806,7 @@ impl SearchAggregator {
             attributes_to_search_on_total_number_of_uses,
             max_terms_number,
             max_vector_size,
+            retrieve_vectors,
             matching_strategy,
             max_limit,
             max_offset,
@@ -873,6 +877,7 @@ impl SearchAggregator {
 
         // vector
         self.max_vector_size = self.max_vector_size.max(max_vector_size);
+        self.retrieve_vectors |= retrieve_vectors;
         self.semantic_ratio |= semantic_ratio;
         self.hybrid |= hybrid;
         self.embedder |= embedder;
@@ -929,6 +934,7 @@ impl SearchAggregator {
             attributes_to_search_on_total_number_of_uses,
             max_terms_number,
             max_vector_size,
+            retrieve_vectors,
             matching_strategy,
             max_limit,
             max_offset,
@@ -991,6 +997,7 @@ impl SearchAggregator {
                 },
                 "vector": {
                     "max_vector_size": max_vector_size,
+                    "retrieve_vectors": retrieve_vectors,
                 },
                 "hybrid": {
                     "enabled": hybrid,
@@ -1079,6 +1086,7 @@ impl MultiSearchAggregator {
                     page: _,
                     hits_per_page: _,
                     attributes_to_retrieve: _,
+                    retrieve_vectors: _,
                     attributes_to_crop: _,
                     crop_length: _,
                     attributes_to_highlight: _,
@@ -1534,6 +1542,9 @@ pub struct DocumentsFetchAggregator {
     // if a filter was used
     per_filter: bool,
 
+    #[serde(rename = "vector.retrieve_vectors")]
+    retrieve_vectors: bool,
+
     // pagination
     #[serde(rename = "pagination.max_limit")]
     max_limit: usize,
@@ -1543,18 +1554,21 @@ pub struct DocumentsFetchAggregator {
 
 impl DocumentsFetchAggregator {
     pub fn from_query(query: &DocumentFetchKind, request: &HttpRequest) -> Self {
-        let (limit, offset) = match query {
-            DocumentFetchKind::PerDocumentId => (1, 0),
-            DocumentFetchKind::Normal { limit, offset, .. } => (*limit, *offset),
+        let (limit, offset, retrieve_vectors) = match query {
+            DocumentFetchKind::PerDocumentId { retrieve_vectors } => (1, 0, *retrieve_vectors),
+            DocumentFetchKind::Normal { limit, offset, retrieve_vectors, .. } => {
+                (*limit, *offset, *retrieve_vectors)
+            }
         };
         Self {
             timestamp: Some(OffsetDateTime::now_utc()),
             user_agents: extract_user_agents(request).into_iter().collect(),
             total_received: 1,
-            per_document_id: matches!(query, DocumentFetchKind::PerDocumentId),
+            per_document_id: matches!(query, DocumentFetchKind::PerDocumentId { .. }),
             per_filter: matches!(query, DocumentFetchKind::Normal { with_filter, .. } if *with_filter),
             max_limit: limit,
             max_offset: offset,
+            retrieve_vectors,
         }
     }
 
@@ -1568,6 +1582,7 @@ impl DocumentsFetchAggregator {
             per_filter,
             max_limit,
             max_offset,
+            retrieve_vectors,
         } = other;
 
         if self.timestamp.is_none() {
@@ -1583,6 +1598,8 @@ impl DocumentsFetchAggregator {
 
         self.max_limit = self.max_limit.max(max_limit);
         self.max_offset = self.max_offset.max(max_offset);
+
+        self.retrieve_vectors |= retrieve_vectors;
     }
 
     pub fn into_event(self, user: &User, event_name: &str) -> Option<Track> {
@@ -1623,6 +1640,7 @@ pub struct SimilarAggregator {
 
     // Whether a non-default embedder was specified
     embedder: bool,
+    retrieve_vectors: bool,
 
     // pagination
     max_limit: usize,
@@ -1646,6 +1664,7 @@ impl SimilarAggregator {
             offset,
             limit,
             attributes_to_retrieve: _,
+            retrieve_vectors,
             show_ranking_score,
             show_ranking_score_details,
             filter,
@@ -1690,6 +1709,7 @@ impl SimilarAggregator {
         ret.ranking_score_threshold = ranking_score_threshold.is_some();
 
         ret.embedder = embedder.is_some();
+        ret.retrieve_vectors = *retrieve_vectors;
 
         ret
     }
@@ -1722,6 +1742,7 @@ impl SimilarAggregator {
             show_ranking_score_details,
             embedder,
             ranking_score_threshold,
+            retrieve_vectors,
         } = other;
 
         if self.timestamp.is_none() {
@@ -1751,6 +1772,7 @@ impl SimilarAggregator {
         }
 
         self.embedder |= embedder;
+        self.retrieve_vectors |= retrieve_vectors;
 
         // pagination
         self.max_limit = self.max_limit.max(max_limit);
@@ -1785,6 +1807,7 @@ impl SimilarAggregator {
             show_ranking_score_details,
             embedder,
             ranking_score_threshold,
+            retrieve_vectors,
         } = self;
 
         if total_received == 0 {
@@ -1811,6 +1834,9 @@ impl SimilarAggregator {
                    "avg_criteria_number": format!("{:.2}", filter_sum_of_criteria_terms as f64 / filter_total_number_of_criteria as f64),
                    "most_used_syntax": used_syntax.iter().max_by_key(|(_, v)| *v).map(|(k, _)| json!(k)).unwrap_or_else(|| json!(null)),
                 },
+                "vector": {
+                    "retrieve_vectors": retrieve_vectors,
+                },
                 "hybrid": {
                     "embedder": embedder,
                 },

meilisearch/src/routes/indexes/documents.rs

@@ -16,6 +16,7 @@ use meilisearch_types::error::{Code, ResponseError};
 use meilisearch_types::heed::RoTxn;
 use meilisearch_types::index_uid::IndexUid;
 use meilisearch_types::milli::update::IndexDocumentsMethod;
+use meilisearch_types::milli::vector::parsed_vectors::ExplicitVectors;
 use meilisearch_types::milli::DocumentId;
 use meilisearch_types::star_or::OptionStarOrList;
 use meilisearch_types::tasks::KindWithContent;
@@ -39,7 +40,7 @@ use crate::extractors::sequential_extractor::SeqHandler;
 use crate::routes::{
     get_task_id, is_dry_run, PaginationView, SummarizedTaskView, PAGINATION_DEFAULT_LIMIT,
 };
-use crate::search::parse_filter;
+use crate::search::{parse_filter, RetrieveVectors};
 use crate::Opt;
 
 static ACCEPTED_CONTENT_TYPE: Lazy<Vec<String>> = Lazy::new(|| {
@@ -94,6 +95,8 @@ pub fn configure(cfg: &mut web::ServiceConfig) {
 pub struct GetDocument {
     #[deserr(default, error = DeserrQueryParamError<InvalidDocumentFields>)]
     fields: OptionStarOrList<String>,
+    #[deserr(default, error = DeserrQueryParamError<InvalidDocumentRetrieveVectors>)]
+    retrieve_vectors: Param<bool>,
 }
 
 pub async fn get_document(
@@ -107,13 +110,20 @@ pub async fn get_document(
     debug!(parameters = ?params, "Get document");
     let index_uid = IndexUid::try_from(index_uid)?;
 
-    analytics.get_fetch_documents(&DocumentFetchKind::PerDocumentId, &req);
-
-    let GetDocument { fields } = params.into_inner();
+    let GetDocument { fields, retrieve_vectors: param_retrieve_vectors } = params.into_inner();
     let attributes_to_retrieve = fields.merge_star_and_none();
 
+    let features = index_scheduler.features();
+    let retrieve_vectors = RetrieveVectors::new(param_retrieve_vectors.0, features)?;
+
+    analytics.get_fetch_documents(
+        &DocumentFetchKind::PerDocumentId { retrieve_vectors: param_retrieve_vectors.0 },
+        &req,
+    );
+
     let index = index_scheduler.index(&index_uid)?;
-    let document = retrieve_document(&index, &document_id, attributes_to_retrieve)?;
+    let document =
+        retrieve_document(&index, &document_id, attributes_to_retrieve, retrieve_vectors)?;
     debug!(returns = ?document, "Get document");
     Ok(HttpResponse::Ok().json(document))
 }
@@ -153,6 +163,8 @@ pub struct BrowseQueryGet {
     limit: Param<usize>,
     #[deserr(default, error = DeserrQueryParamError<InvalidDocumentFields>)]
     fields: OptionStarOrList<String>,
+    #[deserr(default, error = DeserrQueryParamError<InvalidDocumentRetrieveVectors>)]
+    retrieve_vectors: Param<bool>,
     #[deserr(default, error = DeserrQueryParamError<InvalidDocumentFilter>)]
     filter: Option<String>,
 }
@@ -166,6 +178,8 @@ pub struct BrowseQuery {
     limit: usize,
     #[deserr(default, error = DeserrJsonError<InvalidDocumentFields>)]
     fields: Option<Vec<String>>,
+    #[deserr(default, error = DeserrJsonError<InvalidDocumentRetrieveVectors>)]
+    retrieve_vectors: bool,
     #[deserr(default, error = DeserrJsonError<InvalidDocumentFilter>)]
     filter: Option<Value>,
 }
@@ -185,6 +199,7 @@ pub async fn documents_by_query_post(
             with_filter: body.filter.is_some(),
             limit: body.limit,
             offset: body.offset,
+            retrieve_vectors: body.retrieve_vectors,
         },
         &req,
     );
@@ -201,7 +216,7 @@ pub async fn get_documents(
 ) -> Result<HttpResponse, ResponseError> {
     debug!(parameters = ?params, "Get documents GET");
 
-    let BrowseQueryGet { limit, offset, fields, filter } = params.into_inner();
+    let BrowseQueryGet { limit, offset, fields, retrieve_vectors, filter } = params.into_inner();
 
     let filter = match filter {
         Some(f) => match serde_json::from_str(&f) {
@@ -215,6 +230,7 @@ pub async fn get_documents(
         offset: offset.0,
         limit: limit.0,
         fields: fields.merge_star_and_none(),
+        retrieve_vectors: retrieve_vectors.0,
         filter,
     };
 
@@ -223,6 +239,7 @@ pub async fn get_documents(
             with_filter: query.filter.is_some(),
             limit: query.limit,
             offset: query.offset,
+            retrieve_vectors: query.retrieve_vectors,
         },
         &req,
     );
@@ -236,10 +253,14 @@ fn documents_by_query(
     query: BrowseQuery,
 ) -> Result<HttpResponse, ResponseError> {
     let index_uid = IndexUid::try_from(index_uid.into_inner())?;
-    let BrowseQuery { offset, limit, fields, filter } = query;
+    let BrowseQuery { offset, limit, fields, retrieve_vectors, filter } = query;
+
+    let features = index_scheduler.features();
+    let retrieve_vectors = RetrieveVectors::new(retrieve_vectors, features)?;
 
     let index = index_scheduler.index(&index_uid)?;
-    let (total, documents) = retrieve_documents(&index, offset, limit, filter, fields)?;
+    let (total, documents) =
+        retrieve_documents(&index, offset, limit, filter, fields, retrieve_vectors)?;
 
     let ret = PaginationView::new(offset, limit, total as usize, documents);
 
@@ -579,13 +600,44 @@ fn some_documents<'a, 't: 'a>(
     index: &'a Index,
     rtxn: &'t RoTxn,
     doc_ids: impl IntoIterator<Item = DocumentId> + 'a,
+    retrieve_vectors: RetrieveVectors,
 ) -> Result<impl Iterator<Item = Result<Document, ResponseError>> + 'a, ResponseError> {
     let fields_ids_map = index.fields_ids_map(rtxn)?;
     let all_fields: Vec<_> = fields_ids_map.iter().map(|(id, _)| id).collect();
+    let embedding_configs = index.embedding_configs(rtxn)?;
 
     Ok(index.iter_documents(rtxn, doc_ids)?.map(move |ret| {
-        ret.map_err(ResponseError::from).and_then(|(_key, document)| -> Result<_, ResponseError> {
-            Ok(milli::obkv_to_json(&all_fields, &fields_ids_map, document)?)
+        ret.map_err(ResponseError::from).and_then(|(key, document)| -> Result<_, ResponseError> {
+            let mut document = milli::obkv_to_json(&all_fields, &fields_ids_map, document)?;
+            match retrieve_vectors {
+                RetrieveVectors::Ignore => {}
+                RetrieveVectors::Hide => {
+                    document.remove("_vectors");
+                }
+                RetrieveVectors::Retrieve => {
+                    let mut vectors = match document.remove("_vectors") {
+                        Some(Value::Object(map)) => map,
+                        _ => Default::default(),
+                    };
+                    for (name, vector) in index.embeddings(rtxn, key)? {
+                        let user_provided = embedding_configs
+                            .iter()
+                            .find(|conf| conf.name == name)
+                            .is_some_and(|conf| conf.user_provided.contains(key));
+                        let embeddings = ExplicitVectors {
+                            embeddings: Some(vector.into()),
+                            regenerate: !user_provided,
+                        };
+                        vectors.insert(
+                            name,
+                            serde_json::to_value(embeddings).map_err(MeilisearchHttpError::from)?,
+                        );
+                    }
+                    document.insert("_vectors".into(), vectors.into());
+                }
+            }
+
+            Ok(document)
         })
     }))
 }
@@ -596,6 +648,7 @@ fn retrieve_documents<S: AsRef<str>>(
     limit: usize,
     filter: Option<Value>,
     attributes_to_retrieve: Option<Vec<S>>,
+    retrieve_vectors: RetrieveVectors,
 ) -> Result<(u64, Vec<Document>), ResponseError> {
     let rtxn = index.read_txn()?;
     let filter = &filter;
@@ -620,53 +673,57 @@ fn retrieve_documents<S: AsRef<str>>(
     let (it, number_of_documents) = {
         let number_of_documents = candidates.len();
         (
-            some_documents(index, &rtxn, candidates.into_iter().skip(offset).take(limit))?,
+            some_documents(
+                index,
+                &rtxn,
+                candidates.into_iter().skip(offset).take(limit),
+                retrieve_vectors,
+            )?,
             number_of_documents,
         )
     };
 
-    let documents: Result<Vec<_>, ResponseError> = it
+    let documents: Vec<_> = it
         .map(|document| {
             Ok(match &attributes_to_retrieve {
                 Some(attributes_to_retrieve) => permissive_json_pointer::select_values(
                     &document?,
-                    attributes_to_retrieve.iter().map(|s| s.as_ref()),
+                    attributes_to_retrieve.iter().map(|s| s.as_ref()).chain(
+                        (retrieve_vectors == RetrieveVectors::Retrieve).then_some("_vectors"),
+                    ),
                 ),
                 None => document?,
             })
         })
-        .collect();
+        .collect::<Result<_, ResponseError>>()?;
 
-    Ok((number_of_documents, documents?))
+    Ok((number_of_documents, documents))
 }
 
 fn retrieve_document<S: AsRef<str>>(
     index: &Index,
     doc_id: &str,
     attributes_to_retrieve: Option<Vec<S>>,
+    retrieve_vectors: RetrieveVectors,
 ) -> Result<Document, ResponseError> {
     let txn = index.read_txn()?;
 
-    let fields_ids_map = index.fields_ids_map(&txn)?;
-    let all_fields: Vec<_> = fields_ids_map.iter().map(|(id, _)| id).collect();
-
     let internal_id = index
         .external_documents_ids()
         .get(&txn, doc_id)?
         .ok_or_else(|| MeilisearchHttpError::DocumentNotFound(doc_id.to_string()))?;
 
-    let document = index
-        .documents(&txn, std::iter::once(internal_id))?
-        .into_iter()
+    let document = some_documents(index, &txn, Some(internal_id), retrieve_vectors)?
         .next()
-        .map(|(_, d)| d)
-        .ok_or_else(|| MeilisearchHttpError::DocumentNotFound(doc_id.to_string()))?;
+        .ok_or_else(|| MeilisearchHttpError::DocumentNotFound(doc_id.to_string()))??;
 
-    let document = meilisearch_types::milli::obkv_to_json(&all_fields, &fields_ids_map, document)?;
     let document = match &attributes_to_retrieve {
         Some(attributes_to_retrieve) => permissive_json_pointer::select_values(
             &document,
-            attributes_to_retrieve.iter().map(|s| s.as_ref()),
+            attributes_to_retrieve
+                .iter()
+                .map(|s| s.as_ref())
+                .chain((retrieve_vectors == RetrieveVectors::Retrieve).then_some("_vectors")),
         ),
         None => document,
     };

meilisearch/src/routes/indexes/facet_search.rs

@@ -115,6 +115,7 @@ impl From<FacetSearchQuery> for SearchQuery {
             page: None,
             hits_per_page: None,
             attributes_to_retrieve: None,
+            retrieve_vectors: false,
             attributes_to_crop: None,
             crop_length: DEFAULT_CROP_LENGTH(),
             attributes_to_highlight: None,

meilisearch/src/routes/indexes/search.rs

@@ -20,9 +20,9 @@ use crate::extractors::sequential_extractor::SeqHandler;
 use crate::metrics::MEILISEARCH_DEGRADED_SEARCH_REQUESTS;
 use crate::search::{
     add_search_rules, perform_search, HybridQuery, MatchingStrategy, RankingScoreThreshold,
-    SearchKind, SearchQuery, SemanticRatio, DEFAULT_CROP_LENGTH, DEFAULT_CROP_MARKER,
-    DEFAULT_HIGHLIGHT_POST_TAG, DEFAULT_HIGHLIGHT_PRE_TAG, DEFAULT_SEARCH_LIMIT,
-    DEFAULT_SEARCH_OFFSET, DEFAULT_SEMANTIC_RATIO,
+    RetrieveVectors, SearchKind, SearchQuery, SemanticRatio, DEFAULT_CROP_LENGTH,
+    DEFAULT_CROP_MARKER, DEFAULT_HIGHLIGHT_POST_TAG, DEFAULT_HIGHLIGHT_PRE_TAG,
+    DEFAULT_SEARCH_LIMIT, DEFAULT_SEARCH_OFFSET, DEFAULT_SEMANTIC_RATIO,
 };
 use crate::search_queue::SearchQueue;
 
@@ -51,6 +51,8 @@ pub struct SearchQueryGet {
     hits_per_page: Option<Param<usize>>,
     #[deserr(default, error = DeserrQueryParamError<InvalidSearchAttributesToRetrieve>)]
     attributes_to_retrieve: Option<CS<String>>,
+    #[deserr(default, error = DeserrQueryParamError<InvalidSearchRetrieveVectors>)]
+    retrieve_vectors: Param<bool>,
     #[deserr(default, error = DeserrQueryParamError<InvalidSearchAttributesToCrop>)]
     attributes_to_crop: Option<CS<String>>,
     #[deserr(default = Param(DEFAULT_CROP_LENGTH()), error = DeserrQueryParamError<InvalidSearchCropLength>)]
@@ -153,6 +155,7 @@ impl From<SearchQueryGet> for SearchQuery {
             page: other.page.as_deref().copied(),
             hits_per_page: other.hits_per_page.as_deref().copied(),
             attributes_to_retrieve: other.attributes_to_retrieve.map(|o| o.into_iter().collect()),
+            retrieve_vectors: other.retrieve_vectors.0,
             attributes_to_crop: other.attributes_to_crop.map(|o| o.into_iter().collect()),
             crop_length: other.crop_length.0,
             attributes_to_highlight: other.attributes_to_highlight.map(|o| o.into_iter().collect()),
@@ -222,10 +225,12 @@ pub async fn search_with_url_query(
     let features = index_scheduler.features();
 
     let search_kind = search_kind(&query, index_scheduler.get_ref(), &index, features)?;
-
+    let retrieve_vector = RetrieveVectors::new(query.retrieve_vectors, features)?;
     let _permit = search_queue.try_get_search_permit().await?;
-    let search_result =
-        tokio::task::spawn_blocking(move || perform_search(&index, query, search_kind)).await?;
+    let search_result = tokio::task::spawn_blocking(move || {
+        perform_search(&index, query, search_kind, retrieve_vector)
+    })
+    .await?;
     if let Ok(ref search_result) = search_result {
         aggregate.succeed(search_result);
     }
@@ -262,10 +267,13 @@ pub async fn search_with_post(
     let features = index_scheduler.features();
 
     let search_kind = search_kind(&query, index_scheduler.get_ref(), &index, features)?;
+    let retrieve_vectors = RetrieveVectors::new(query.retrieve_vectors, features)?;
 
     let _permit = search_queue.try_get_search_permit().await?;
-    let search_result =
-        tokio::task::spawn_blocking(move || perform_search(&index, query, search_kind)).await?;
+    let search_result = tokio::task::spawn_blocking(move || {
+        perform_search(&index, query, search_kind, retrieve_vectors)
+    })
+    .await?;
     if let Ok(ref search_result) = search_result {
         aggregate.succeed(search_result);
         if search_result.degraded {
@@ -287,11 +295,10 @@ pub fn search_kind(
     features: RoFeatures,
 ) -> Result<SearchKind, ResponseError> {
     if query.vector.is_some() {
-        features.check_vector("Passing `vector` as a query parameter")?;
+        features.check_vector("Passing `vector` as a parameter")?;
     }
-
     if query.hybrid.is_some() {
-        features.check_vector("Passing `hybrid` as a query parameter")?;
+        features.check_vector("Passing `hybrid` as a parameter")?;
     }
 
     // regardless of anything, always do a keyword search when we don't have a vector and the query is whitespace or missing

meilisearch/src/routes/indexes/similar.rs

@@ -4,11 +4,7 @@ use deserr::actix_web::{AwebJson, AwebQueryParameter};
 use index_scheduler::IndexScheduler;
 use meilisearch_types::deserr::query_params::Param;
 use meilisearch_types::deserr::{DeserrJsonError, DeserrQueryParamError};
-use meilisearch_types::error::deserr_codes::{
-    InvalidEmbedder, InvalidSimilarAttributesToRetrieve, InvalidSimilarFilter, InvalidSimilarId,
-    InvalidSimilarLimit, InvalidSimilarOffset, InvalidSimilarRankingScoreThreshold,
-    InvalidSimilarShowRankingScore, InvalidSimilarShowRankingScoreDetails,
-};
+use meilisearch_types::error::deserr_codes::*;
 use meilisearch_types::error::{ErrorCode as _, ResponseError};
 use meilisearch_types::index_uid::IndexUid;
 use meilisearch_types::keys::actions;
@@ -21,8 +17,8 @@ use crate::analytics::{Analytics, SimilarAggregator};
 use crate::extractors::authentication::GuardedData;
 use crate::extractors::sequential_extractor::SeqHandler;
 use crate::search::{
-    add_search_rules, perform_similar, RankingScoreThresholdSimilar, SearchKind, SimilarQuery,
-    SimilarResult, DEFAULT_SEARCH_LIMIT, DEFAULT_SEARCH_OFFSET,
+    add_search_rules, perform_similar, RankingScoreThresholdSimilar, RetrieveVectors, SearchKind,
+    SimilarQuery, SimilarResult, DEFAULT_SEARCH_LIMIT, DEFAULT_SEARCH_OFFSET,
 };
 
 pub fn configure(cfg: &mut web::ServiceConfig) {
@@ -97,6 +93,8 @@ async fn similar(
 
     features.check_vector("Using the similar API")?;
 
+    let retrieve_vectors = RetrieveVectors::new(query.retrieve_vectors, features)?;
+
     // Tenant token search_rules.
     if let Some(search_rules) = index_scheduler.filters().get_index_search_rules(&index_uid) {
         add_search_rules(&mut query.filter, search_rules);
@@ -107,8 +105,10 @@ async fn similar(
     let (embedder_name, embedder) =
         SearchKind::embedder(&index_scheduler, &index, query.embedder.as_deref(), None)?;
 
-    tokio::task::spawn_blocking(move || perform_similar(&index, query, embedder_name, embedder))
-        .await?
+    tokio::task::spawn_blocking(move || {
+        perform_similar(&index, query, embedder_name, embedder, retrieve_vectors)
+    })
+    .await?
 }
 
 #[derive(Debug, deserr::Deserr)]
@@ -122,6 +122,8 @@ pub struct SimilarQueryGet {
     limit: Param<usize>,
     #[deserr(default, error = DeserrQueryParamError<InvalidSimilarAttributesToRetrieve>)]
     attributes_to_retrieve: Option<CS<String>>,
+    #[deserr(default, error = DeserrQueryParamError<InvalidSimilarRetrieveVectors>)]
+    retrieve_vectors: Param<bool>,
     #[deserr(default, error = DeserrQueryParamError<InvalidSimilarFilter>)]
     filter: Option<String>,
     #[deserr(default, error = DeserrQueryParamError<InvalidSimilarShowRankingScore>)]
@@ -156,6 +158,7 @@ impl TryFrom<SimilarQueryGet> for SimilarQuery {
             offset,
             limit,
             attributes_to_retrieve,
+            retrieve_vectors,
             filter,
             show_ranking_score,
             show_ranking_score_details,
@@ -180,6 +183,7 @@ impl TryFrom<SimilarQueryGet> for SimilarQuery {
             filter,
             embedder,
             attributes_to_retrieve: attributes_to_retrieve.map(|o| o.into_iter().collect()),
+            retrieve_vectors: retrieve_vectors.0,
             show_ranking_score: show_ranking_score.0,
             show_ranking_score_details: show_ranking_score_details.0,
             ranking_score_threshold: ranking_score_threshold.map(|x| x.0),

meilisearch/src/routes/multi_search.rs

@@ -15,7 +15,7 @@ use crate::extractors::authentication::{AuthenticationError, GuardedData};
 use crate::extractors::sequential_extractor::SeqHandler;
 use crate::routes::indexes::search::search_kind;
 use crate::search::{
-    add_search_rules, perform_search, SearchQueryWithIndex, SearchResultWithIndex,
+    add_search_rules, perform_search, RetrieveVectors, SearchQueryWithIndex, SearchResultWithIndex,
 };
 use crate::search_queue::SearchQueue;
 
@@ -83,11 +83,14 @@ pub async fn multi_search_with_post(
 
             let search_kind = search_kind(&query, index_scheduler.get_ref(), &index, features)
                 .with_index(query_index)?;
+            let retrieve_vector =
+                RetrieveVectors::new(query.retrieve_vectors, features).with_index(query_index)?;
 
-            let search_result =
-                tokio::task::spawn_blocking(move || perform_search(&index, query, search_kind))
-                    .await
-                    .with_index(query_index)?;
+            let search_result = tokio::task::spawn_blocking(move || {
+                perform_search(&index, query, search_kind, retrieve_vector)
+            })
+            .await
+            .with_index(query_index)?;
 
             search_results.push(SearchResultWithIndex {
                 index_uid: index_uid.into_inner(),

meilisearch/src/search.rs

@@ -15,6 +15,7 @@ use meilisearch_types::error::{Code, ResponseError};
 use meilisearch_types::heed::RoTxn;
 use meilisearch_types::index_uid::IndexUid;
 use meilisearch_types::milli::score_details::{ScoreDetails, ScoringStrategy};
+use meilisearch_types::milli::vector::parsed_vectors::ExplicitVectors;
 use meilisearch_types::milli::vector::Embedder;
 use meilisearch_types::milli::{FacetValueHit, OrderBy, SearchForFacetValues, TimeBudget};
 use meilisearch_types::settings::DEFAULT_PAGINATION_MAX_TOTAL_HITS;
@@ -59,6 +60,8 @@ pub struct SearchQuery {
     pub hits_per_page: Option<usize>,
     #[deserr(default, error = DeserrJsonError<InvalidSearchAttributesToRetrieve>)]
     pub attributes_to_retrieve: Option<BTreeSet<String>>,
+    #[deserr(default, error = DeserrJsonError<InvalidSearchRetrieveVectors>)]
+    pub retrieve_vectors: bool,
     #[deserr(default, error = DeserrJsonError<InvalidSearchAttributesToCrop>)]
     pub attributes_to_crop: Option<Vec<String>>,
     #[deserr(default, error = DeserrJsonError<InvalidSearchCropLength>, default = DEFAULT_CROP_LENGTH())]
@@ -141,6 +144,7 @@ impl fmt::Debug for SearchQuery {
             page,
             hits_per_page,
             attributes_to_retrieve,
+            retrieve_vectors,
             attributes_to_crop,
             crop_length,
             attributes_to_highlight,
@@ -173,6 +177,9 @@ impl fmt::Debug for SearchQuery {
         if let Some(q) = q {
             debug.field("q", &q);
         }
+        if *retrieve_vectors {
+            debug.field("retrieve_vectors", &retrieve_vectors);
+        }
         if let Some(v) = vector {
             if v.len() < 10 {
                 debug.field("vector", &v);
@@ -370,6 +377,8 @@ pub struct SearchQueryWithIndex {
     pub hits_per_page: Option<usize>,
     #[deserr(default, error = DeserrJsonError<InvalidSearchAttributesToRetrieve>)]
     pub attributes_to_retrieve: Option<BTreeSet<String>>,
+    #[deserr(default, error = DeserrJsonError<InvalidSearchRetrieveVectors>)]
+    pub retrieve_vectors: bool,
     #[deserr(default, error = DeserrJsonError<InvalidSearchAttributesToCrop>)]
     pub attributes_to_crop: Option<Vec<String>>,
     #[deserr(default, error = DeserrJsonError<InvalidSearchCropLength>, default = DEFAULT_CROP_LENGTH())]
@@ -413,6 +422,7 @@ impl SearchQueryWithIndex {
             page,
             hits_per_page,
             attributes_to_retrieve,
+            retrieve_vectors,
             attributes_to_crop,
             crop_length,
             attributes_to_highlight,
@@ -440,6 +450,7 @@ impl SearchQueryWithIndex {
                 page,
                 hits_per_page,
                 attributes_to_retrieve,
+                retrieve_vectors,
                 attributes_to_crop,
                 crop_length,
                 attributes_to_highlight,
@@ -478,6 +489,8 @@ pub struct SimilarQuery {
     pub embedder: Option<String>,
     #[deserr(default, error = DeserrJsonError<InvalidSimilarAttributesToRetrieve>)]
     pub attributes_to_retrieve: Option<BTreeSet<String>>,
+    #[deserr(default, error = DeserrJsonError<InvalidSimilarRetrieveVectors>)]
+    pub retrieve_vectors: bool,
     #[deserr(default, error = DeserrJsonError<InvalidSimilarShowRankingScore>, default)]
     pub show_ranking_score: bool,
     #[deserr(default, error = DeserrJsonError<InvalidSimilarShowRankingScoreDetails>, default)]
@@ -810,6 +823,7 @@ pub fn perform_search(
     index: &Index,
     query: SearchQuery,
     search_kind: SearchKind,
+    retrieve_vectors: RetrieveVectors,
 ) -> Result<SearchResult, MeilisearchHttpError> {
     let before_search = Instant::now();
     let rtxn = index.read_txn()?;
@@ -847,6 +861,8 @@ pub fn perform_search(
         page,
         hits_per_page,
         attributes_to_retrieve,
+        // use the enum passed as parameter
+        retrieve_vectors: _,
         attributes_to_crop,
         crop_length,
         attributes_to_highlight,
@@ -870,6 +886,7 @@ pub fn perform_search(
 
     let format = AttributesFormat {
         attributes_to_retrieve,
+        retrieve_vectors,
         attributes_to_highlight,
         attributes_to_crop,
         crop_length,
@@ -953,6 +970,7 @@ pub fn perform_search(
 
 struct AttributesFormat {
     attributes_to_retrieve: Option<BTreeSet<String>>,
+    retrieve_vectors: RetrieveVectors,
     attributes_to_highlight: Option<HashSet<String>>,
     attributes_to_crop: Option<Vec<String>>,
     crop_length: usize,
@@ -965,6 +983,36 @@ struct AttributesFormat {
     show_ranking_score_details: bool,
 }
 
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum RetrieveVectors {
+    /// Do not touch the `_vectors` field
+    ///
+    /// this is the behavior when the vectorStore feature is disabled
+    Ignore,
+    /// Remove the `_vectors` field
+    ///
+    /// this is the behavior when the vectorStore feature is enabled, and `retrieveVectors` is `false`
+    Hide,
+    /// Retrieve vectors from the DB and merge them into the `_vectors` field
+    ///
+    /// this is the behavior when the vectorStore feature is enabled, and `retrieveVectors` is `true`
+    Retrieve,
+}
+
+impl RetrieveVectors {
+    pub fn new(
+        retrieve_vector: bool,
+        features: index_scheduler::RoFeatures,
+    ) -> Result<Self, index_scheduler::Error> {
+        match (retrieve_vector, features.check_vector("Passing `retrieveVectors` as a parameter")) {
+            (true, Ok(())) => Ok(Self::Retrieve),
+            (true, Err(error)) => Err(error),
+            (false, Ok(())) => Ok(Self::Hide),
+            (false, Err(_)) => Ok(Self::Ignore),
+        }
+    }
+}
+
 fn make_hits(
     index: &Index,
     rtxn: &RoTxn<'_>,
@@ -974,10 +1022,32 @@ fn make_hits(
     document_scores: Vec<Vec<ScoreDetails>>,
 ) -> Result<Vec<SearchHit>, MeilisearchHttpError> {
     let fields_ids_map = index.fields_ids_map(rtxn).unwrap();
-    let displayed_ids = index
-        .displayed_fields_ids(rtxn)?
-        .map(|fields| fields.into_iter().collect::<BTreeSet<_>>())
-        .unwrap_or_else(|| fields_ids_map.iter().map(|(id, _)| id).collect());
+    let displayed_ids =
+        index.displayed_fields_ids(rtxn)?.map(|fields| fields.into_iter().collect::<BTreeSet<_>>());
+
+    let vectors_fid = fields_ids_map.id(milli::vector::parsed_vectors::RESERVED_VECTORS_FIELD_NAME);
+
+    let vectors_is_hidden = match (&displayed_ids, vectors_fid) {
+        // displayed_ids is a wildcard, so `_vectors` can be displayed regardless of its fid
+        (None, _) => false,
+        // displayed_ids is a finite list, and `_vectors` cannot be part of it because it is not an existing field
+        (Some(_), None) => true,
+        // displayed_ids is a finit list, so hide if `_vectors` is not part of it
+        (Some(map), Some(vectors_fid)) => map.contains(&vectors_fid),
+    };
+
+    let retrieve_vectors = if let RetrieveVectors::Retrieve = format.retrieve_vectors {
+        if vectors_is_hidden {
+            RetrieveVectors::Hide
+        } else {
+            RetrieveVectors::Retrieve
+        }
+    } else {
+        format.retrieve_vectors
+    };
+
+    let displayed_ids =
+        displayed_ids.unwrap_or_else(|| fields_ids_map.iter().map(|(id, _)| id).collect());
     let fids = |attrs: &BTreeSet<String>| {
         let mut ids = BTreeSet::new();
         for attr in attrs {
@@ -1000,6 +1070,7 @@ fn make_hits(
         .intersection(&displayed_ids)
         .cloned()
         .collect();
+
     let attr_to_highlight = format.attributes_to_highlight.unwrap_or_default();
     let attr_to_crop = format.attributes_to_crop.unwrap_or_default();
     let formatted_options = compute_formatted_options(
@@ -1033,18 +1104,48 @@ fn make_hits(
     formatter_builder.highlight_prefix(format.highlight_pre_tag);
     formatter_builder.highlight_suffix(format.highlight_post_tag);
     let mut documents = Vec::new();
+    let embedding_configs = index.embedding_configs(rtxn)?;
     let documents_iter = index.documents(rtxn, documents_ids)?;
-    for ((_id, obkv), score) in documents_iter.into_iter().zip(document_scores.into_iter()) {
+    for ((id, obkv), score) in documents_iter.into_iter().zip(document_scores.into_iter()) {
         // First generate a document with all the displayed fields
         let displayed_document = make_document(&displayed_ids, &fields_ids_map, obkv)?;
 
+        let add_vectors_fid =
+            vectors_fid.filter(|_fid| retrieve_vectors == RetrieveVectors::Retrieve);
+
         // select the attributes to retrieve
         let attributes_to_retrieve = to_retrieve_ids
             .iter()
+            // skip the vectors_fid if RetrieveVectors::Hide
+            .filter(|fid| match vectors_fid {
+                Some(vectors_fid) => {
+                    !(retrieve_vectors == RetrieveVectors::Hide && **fid == vectors_fid)
+                }
+                None => true,
+            })
+            // need to retrieve the existing `_vectors` field if the `RetrieveVectors::Retrieve`
+            .chain(add_vectors_fid.iter())
             .map(|&fid| fields_ids_map.name(fid).expect("Missing field name"));
         let mut document =
             permissive_json_pointer::select_values(&displayed_document, attributes_to_retrieve);
 
+        if retrieve_vectors == RetrieveVectors::Retrieve {
+            let mut vectors = match document.remove("_vectors") {
+                Some(Value::Object(map)) => map,
+                _ => Default::default(),
+            };
+            for (name, vector) in index.embeddings(rtxn, id)? {
+                let user_provided = embedding_configs
+                    .iter()
+                    .find(|conf| conf.name == name)
+                    .is_some_and(|conf| conf.user_provided.contains(id));
+                let embeddings =
+                    ExplicitVectors { embeddings: Some(vector.into()), regenerate: !user_provided };
+                vectors.insert(name, serde_json::to_value(embeddings)?);
+            }
+            document.insert("_vectors".into(), vectors.into());
+        }
+
         let (matches_position, formatted) = format_fields(
             &displayed_document,
             &fields_ids_map,
@@ -1114,6 +1215,7 @@ pub fn perform_similar(
     query: SimilarQuery,
     embedder_name: String,
     embedder: Arc<Embedder>,
+    retrieve_vectors: RetrieveVectors,
 ) -> Result<SimilarResult, ResponseError> {
     let before_search = Instant::now();
     let rtxn = index.read_txn()?;
@@ -1125,6 +1227,7 @@ pub fn perform_similar(
         filter: _,
         embedder: _,
         attributes_to_retrieve,
+        retrieve_vectors: _,
         show_ranking_score,
         show_ranking_score_details,
         ranking_score_threshold,
@@ -1171,6 +1274,7 @@ pub fn perform_similar(
 
     let format = AttributesFormat {
         attributes_to_retrieve,
+        retrieve_vectors,
         attributes_to_highlight: None,
         attributes_to_crop: None,
         crop_length: DEFAULT_CROP_LENGTH(),

meilisearch/tests/common/index.rs

@@ -182,14 +182,10 @@ impl Index<'_> {
         self.service.get(url).await
     }
 
-    pub async fn get_document(
-        &self,
-        id: u64,
-        options: Option<GetDocumentOptions>,
-    ) -> (Value, StatusCode) {
+    pub async fn get_document(&self, id: u64, options: Option<Value>) -> (Value, StatusCode) {
         let mut url = format!("/indexes/{}/documents/{}", urlencode(self.uid.as_ref()), id);
-        if let Some(fields) = options.and_then(|o| o.fields) {
-            let _ = write!(url, "?fields={}", fields.join(","));
+        if let Some(options) = options {
+            write!(url, "?{}", yaup::to_string(&options).unwrap()).unwrap();
         }
         self.service.get(url).await
     }
@@ -205,18 +201,11 @@ impl Index<'_> {
     }
 
     pub async fn get_all_documents(&self, options: GetAllDocumentsOptions) -> (Value, StatusCode) {
-        let mut url = format!("/indexes/{}/documents?", urlencode(self.uid.as_ref()));
-        if let Some(limit) = options.limit {
-            let _ = write!(url, "limit={}&", limit);
-        }
-
-        if let Some(offset) = options.offset {
-            let _ = write!(url, "offset={}&", offset);
-        }
-
-        if let Some(attributes_to_retrieve) = options.attributes_to_retrieve {
-            let _ = write!(url, "fields={}&", attributes_to_retrieve.join(","));
-        }
+        let url = format!(
+            "/indexes/{}/documents?{}",
+            urlencode(self.uid.as_ref()),
+            yaup::to_string(&options).unwrap()
+        );
 
         self.service.get(url).await
     }
@@ -435,13 +424,11 @@ impl Index<'_> {
     }
 }
 
-pub struct GetDocumentOptions {
-    pub fields: Option<Vec<&'static str>>,
-}
-
-#[derive(Debug, Default)]
+#[derive(Debug, Default, serde::Serialize)]
+#[serde(rename_all = "camelCase")]
 pub struct GetAllDocumentsOptions {
     pub limit: Option<usize>,
     pub offset: Option<usize>,
-    pub attributes_to_retrieve: Option<Vec<&'static str>>,
+    pub retrieve_vectors: bool,
+    pub fields: Option<Vec<&'static str>>,
 }

meilisearch/tests/common/mod.rs

@@ -6,7 +6,7 @@ pub mod service;
 use std::fmt::{self, Display};
 
 #[allow(unused)]
-pub use index::{GetAllDocumentsOptions, GetDocumentOptions};
+pub use index::GetAllDocumentsOptions;
 use meili_snap::json_string;
 use serde::{Deserialize, Serialize};
 #[allow(unused)]

meilisearch/tests/documents/errors.rs

@@ -795,3 +795,70 @@ async fn fetch_document_by_filter() {
     }
     "###);
 }
+
+#[actix_rt::test]
+async fn retrieve_vectors() {
+    let server = Server::new().await;
+    let index = server.index("doggo");
+
+    // GET ALL DOCUMENTS BY QUERY
+    let (response, _code) = index.get_all_documents_raw("?retrieveVectors=tamo").await;
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Invalid value in parameter `retrieveVectors`: could not parse `tamo` as a boolean, expected either `true` or `false`",
+      "code": "invalid_document_retrieve_vectors",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_document_retrieve_vectors"
+    }
+    "###);
+    let (response, _code) = index.get_all_documents_raw("?retrieveVectors=true").await;
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Passing `retrieveVectors` as a parameter requires enabling the `vector store` experimental feature. See https://github.com/meilisearch/product/discussions/677",
+      "code": "feature_not_enabled",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#feature_not_enabled"
+    }
+    "###);
+
+    // FETCH ALL DOCUMENTS BY POST
+    let (response, _code) =
+        index.get_document_by_filter(json!({ "retrieveVectors": "tamo" })).await;
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Invalid value type at `.retrieveVectors`: expected a boolean, but found a string: `\"tamo\"`",
+      "code": "invalid_document_retrieve_vectors",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_document_retrieve_vectors"
+    }
+    "###);
+    let (response, _code) = index.get_document_by_filter(json!({ "retrieveVectors": true })).await;
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Passing `retrieveVectors` as a parameter requires enabling the `vector store` experimental feature. See https://github.com/meilisearch/product/discussions/677",
+      "code": "feature_not_enabled",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#feature_not_enabled"
+    }
+    "###);
+
+    // GET A SINGLE DOCUMENT
+    let (response, _code) = index.get_document(0, Some(json!({"retrieveVectors": "tamo"}))).await;
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Invalid value in parameter `retrieveVectors`: could not parse `tamo` as a boolean, expected either `true` or `false`",
+      "code": "invalid_document_retrieve_vectors",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_document_retrieve_vectors"
+    }
+    "###);
+    let (response, _code) = index.get_document(0, Some(json!({"retrieveVectors": true}))).await;
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Passing `retrieveVectors` as a parameter requires enabling the `vector store` experimental feature. See https://github.com/meilisearch/product/discussions/677",
+      "code": "feature_not_enabled",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#feature_not_enabled"
+    }
+    "###);
+}

meilisearch/tests/documents/get_documents.rs

@@ -4,7 +4,7 @@ use meili_snap::*;
 use urlencoding::encode as urlencode;
 
 use crate::common::encoder::Encoder;
-use crate::common::{GetAllDocumentsOptions, GetDocumentOptions, Server, Value};
+use crate::common::{GetAllDocumentsOptions, Server, Value};
 use crate::json;
 
 // TODO: partial test since we are testing error, amd error is not yet fully implemented in
@@ -59,8 +59,7 @@ async fn get_document() {
         })
     );
 
-    let (response, code) =
-        index.get_document(0, Some(GetDocumentOptions { fields: Some(vec!["id"]) })).await;
+    let (response, code) = index.get_document(0, Some(json!({ "fields": ["id"] }))).await;
     assert_eq!(code, 200);
     assert_eq!(
         response,
@@ -69,9 +68,8 @@ async fn get_document() {
         })
     );
 
-    let (response, code) = index
-        .get_document(0, Some(GetDocumentOptions { fields: Some(vec!["nested.content"]) }))
-        .await;
+    let (response, code) =
+        index.get_document(0, Some(json!({ "fields": ["nested.content"] }))).await;
     assert_eq!(code, 200);
     assert_eq!(
         response,
@@ -211,7 +209,7 @@ async fn test_get_all_documents_attributes_to_retrieve() {
 
     let (response, code) = index
         .get_all_documents(GetAllDocumentsOptions {
-            attributes_to_retrieve: Some(vec!["name"]),
+            fields: Some(vec!["name"]),
             ..Default::default()
         })
         .await;
@@ -225,9 +223,19 @@ async fn test_get_all_documents_attributes_to_retrieve() {
     assert_eq!(response["limit"], json!(20));
     assert_eq!(response["total"], json!(77));
 
+    let (response, code) = index.get_all_documents_raw("?fields=").await;
+    assert_eq!(code, 200);
+    assert_eq!(response["results"].as_array().unwrap().len(), 20);
+    for results in response["results"].as_array().unwrap() {
+        assert_eq!(results.as_object().unwrap().keys().count(), 0);
+    }
+    assert_eq!(response["offset"], json!(0));
+    assert_eq!(response["limit"], json!(20));
+    assert_eq!(response["total"], json!(77));
+
     let (response, code) = index
         .get_all_documents(GetAllDocumentsOptions {
-            attributes_to_retrieve: Some(vec![]),
+            fields: Some(vec!["wrong"]),
             ..Default::default()
         })
         .await;
@@ -242,22 +250,7 @@ async fn test_get_all_documents_attributes_to_retrieve() {
 
     let (response, code) = index
         .get_all_documents(GetAllDocumentsOptions {
-            attributes_to_retrieve: Some(vec!["wrong"]),
-            ..Default::default()
-        })
-        .await;
-    assert_eq!(code, 200);
-    assert_eq!(response["results"].as_array().unwrap().len(), 20);
-    for results in response["results"].as_array().unwrap() {
-        assert_eq!(results.as_object().unwrap().keys().count(), 0);
-    }
-    assert_eq!(response["offset"], json!(0));
-    assert_eq!(response["limit"], json!(20));
-    assert_eq!(response["total"], json!(77));
-
-    let (response, code) = index
-        .get_all_documents(GetAllDocumentsOptions {
-            attributes_to_retrieve: Some(vec!["name", "tags"]),
+            fields: Some(vec!["name", "tags"]),
             ..Default::default()
         })
         .await;
@@ -270,10 +263,7 @@ async fn test_get_all_documents_attributes_to_retrieve() {
     }
 
     let (response, code) = index
-        .get_all_documents(GetAllDocumentsOptions {
-            attributes_to_retrieve: Some(vec!["*"]),
-            ..Default::default()
-        })
+        .get_all_documents(GetAllDocumentsOptions { fields: Some(vec!["*"]), ..Default::default() })
         .await;
     assert_eq!(code, 200);
     assert_eq!(response["results"].as_array().unwrap().len(), 20);
@@ -283,7 +273,7 @@ async fn test_get_all_documents_attributes_to_retrieve() {
 
     let (response, code) = index
         .get_all_documents(GetAllDocumentsOptions {
-            attributes_to_retrieve: Some(vec!["*", "wrong"]),
+            fields: Some(vec!["*", "wrong"]),
             ..Default::default()
         })
         .await;
@@ -316,12 +306,10 @@ async fn get_document_s_nested_attributes_to_retrieve() {
     assert_eq!(code, 202);
     index.wait_task(1).await;
 
-    let (response, code) =
-        index.get_document(0, Some(GetDocumentOptions { fields: Some(vec!["content"]) })).await;
+    let (response, code) = index.get_document(0, Some(json!({ "fields": ["content"] }))).await;
     assert_eq!(code, 200);
     assert_eq!(response, json!({}));
-    let (response, code) =
-        index.get_document(1, Some(GetDocumentOptions { fields: Some(vec!["content"]) })).await;
+    let (response, code) = index.get_document(1, Some(json!({ "fields": ["content"] }))).await;
     assert_eq!(code, 200);
     assert_eq!(
         response,
@@ -333,9 +321,7 @@ async fn get_document_s_nested_attributes_to_retrieve() {
         })
     );
 
-    let (response, code) = index
-        .get_document(0, Some(GetDocumentOptions { fields: Some(vec!["content.truc"]) }))
-        .await;
+    let (response, code) = index.get_document(0, Some(json!({ "fields": ["content.truc"] }))).await;
     assert_eq!(code, 200);
     assert_eq!(
         response,
@@ -343,9 +329,7 @@ async fn get_document_s_nested_attributes_to_retrieve() {
             "content.truc": "foobar",
         })
     );
-    let (response, code) = index
-        .get_document(1, Some(GetDocumentOptions { fields: Some(vec!["content.truc"]) }))
-        .await;
+    let (response, code) = index.get_document(1, Some(json!({ "fields": ["content.truc"] }))).await;
     assert_eq!(code, 200);
     assert_eq!(
         response,
@@ -540,3 +524,207 @@ async fn get_document_by_filter() {
     }
     "###);
 }
+
+#[actix_rt::test]
+async fn get_document_with_vectors() {
+    let server = Server::new().await;
+    let index = server.index("doggo");
+    let (value, code) = server.set_features(json!({"vectorStore": true})).await;
+    snapshot!(code, @"200 OK");
+    snapshot!(value, @r###"
+    {
+      "vectorStore": true,
+      "metrics": false,
+      "logsRoute": false
+    }
+    "###);
+
+    let (response, code) = index
+        .update_settings(json!({
+          "embedders": {
+              "manual": {
+                  "source": "userProvided",
+                  "dimensions": 3,
+              }
+          },
+        }))
+        .await;
+    snapshot!(code, @"202 Accepted");
+    server.wait_task(response.uid()).await;
+
+    let documents = json!([
+      {"id": 0, "name": "kefir", "_vectors": { "manual": [0, 0, 0] }},
+      {"id": 1, "name": "echo", "_vectors": { "manual": null }},
+    ]);
+    let (value, code) = index.add_documents(documents, None).await;
+    snapshot!(code, @"202 Accepted");
+    index.wait_task(value.uid()).await;
+
+    // by default you shouldn't see the `_vectors` object
+    let (documents, _code) = index.get_all_documents(Default::default()).await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "results": [
+        {
+          "id": 0,
+          "name": "kefir"
+        },
+        {
+          "id": 1,
+          "name": "echo"
+        }
+      ],
+      "offset": 0,
+      "limit": 20,
+      "total": 2
+    }
+    "###);
+    let (documents, _code) = index.get_document(0, None).await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "id": 0,
+      "name": "kefir"
+    }
+    "###);
+
+    // if we try to retrieve the vectors with the `fields` parameter they
+    // still shouldn't be displayed
+    let (documents, _code) = index
+        .get_all_documents(GetAllDocumentsOptions {
+            fields: Some(vec!["name", "_vectors"]),
+            ..Default::default()
+        })
+        .await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "results": [
+        {
+          "name": "kefir"
+        },
+        {
+          "name": "echo"
+        }
+      ],
+      "offset": 0,
+      "limit": 20,
+      "total": 2
+    }
+    "###);
+    let (documents, _code) =
+        index.get_document(0, Some(json!({"fields": ["name", "_vectors"]}))).await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "name": "kefir"
+    }
+    "###);
+
+    // If we specify the retrieve vectors boolean and nothing else we should get the vectors
+    let (documents, _code) = index
+        .get_all_documents(GetAllDocumentsOptions { retrieve_vectors: true, ..Default::default() })
+        .await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "results": [
+        {
+          "id": 0,
+          "name": "kefir",
+          "_vectors": {
+            "manual": {
+              "embeddings": [
+                [
+                  0.0,
+                  0.0,
+                  0.0
+                ]
+              ],
+              "regenerate": false
+            }
+          }
+        },
+        {
+          "id": 1,
+          "name": "echo",
+          "_vectors": {}
+        }
+      ],
+      "offset": 0,
+      "limit": 20,
+      "total": 2
+    }
+    "###);
+    let (documents, _code) = index.get_document(0, Some(json!({"retrieveVectors": true}))).await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "id": 0,
+      "name": "kefir",
+      "_vectors": {
+        "manual": {
+          "embeddings": [
+            [
+              0.0,
+              0.0,
+              0.0
+            ]
+          ],
+          "regenerate": false
+        }
+      }
+    }
+    "###);
+
+    // If we specify the retrieve vectors boolean and exclude vectors form the `fields` we should still get the vectors
+    let (documents, _code) = index
+        .get_all_documents(GetAllDocumentsOptions {
+            retrieve_vectors: true,
+            fields: Some(vec!["name"]),
+            ..Default::default()
+        })
+        .await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "results": [
+        {
+          "name": "kefir",
+          "_vectors": {
+            "manual": {
+              "embeddings": [
+                [
+                  0.0,
+                  0.0,
+                  0.0
+                ]
+              ],
+              "regenerate": false
+            }
+          }
+        },
+        {
+          "name": "echo",
+          "_vectors": {}
+        }
+      ],
+      "offset": 0,
+      "limit": 20,
+      "total": 2
+    }
+    "###);
+    let (documents, _code) =
+        index.get_document(0, Some(json!({"retrieveVectors": true, "fields": ["name"]}))).await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "name": "kefir",
+      "_vectors": {
+        "manual": {
+          "embeddings": [
+            [
+              0.0,
+              0.0,
+              0.0
+            ]
+          ],
+          "regenerate": false
+        }
+      }
+    }
+    "###);
+}

meilisearch/tests/dumps/mod.rs

@@ -1938,3 +1938,210 @@ async fn import_dump_v6_containing_experimental_features() {
         })
         .await;
 }
+
+// In this test we must generate the dump ourselves to ensure the
+// `user provided` vectors are well set
+#[actix_rt::test]
+#[cfg_attr(target_os = "windows", ignore)]
+async fn generate_and_import_dump_containing_vectors() {
+    let temp = tempfile::tempdir().unwrap();
+    let mut opt = default_settings(temp.path());
+    let server = Server::new_with_options(opt.clone()).await.unwrap();
+    let (code, _) = server.set_features(json!({"vectorStore": true})).await;
+    snapshot!(code, @r###"
+    {
+      "vectorStore": true,
+      "metrics": false,
+      "logsRoute": false
+    }
+    "###);
+    let index = server.index("pets");
+    let (response, code) = index
+        .update_settings(json!(
+        {
+            "embedders": {
+                "doggo_embedder": {
+                    "source": "huggingFace",
+                    "model": "sentence-transformers/all-MiniLM-L6-v2",
+                    "revision": "e4ce9877abf3edfe10b0d82785e83bdcb973e22e",
+                    "documentTemplate": "{{doc.doggo}}",
+                }
+            }
+        }
+        ))
+        .await;
+    snapshot!(code, @"202 Accepted");
+    let response = index.wait_task(response.uid()).await;
+    snapshot!(response);
+    let (response, code) = index
+        .add_documents(
+            json!([
+                {"id": 0, "doggo": "kefir", "_vectors": { "doggo_embedder": vec![0; 384] }},
+                {"id": 1, "doggo": "echo", "_vectors": { "doggo_embedder": { "regenerate": false, "embeddings": vec![1; 384] }}},
+                {"id": 2, "doggo": "intel", "_vectors": { "doggo_embedder": { "regenerate": true, "embeddings": vec![2; 384] }}},
+                {"id": 3, "doggo": "bill", "_vectors": { "doggo_embedder": { "regenerate": true }}},
+                {"id": 4, "doggo": "max" },
+            ]),
+            None,
+        )
+        .await;
+    snapshot!(code, @"202 Accepted");
+    let response = index.wait_task(response.uid()).await;
+    snapshot!(response);
+
+    let (response, code) = server.create_dump().await;
+    snapshot!(code, @"202 Accepted");
+    let response = index.wait_task(response.uid()).await;
+    snapshot!(response["status"], @r###""succeeded""###);
+
+    // ========= We made a dump, now we should clear the DB and try to import our dump
+    drop(server);
+    tokio::fs::remove_dir_all(&opt.db_path).await.unwrap();
+    let dump_name = format!("{}.dump", response["details"]["dumpUid"].as_str().unwrap());
+    let dump_path = opt.dump_dir.join(dump_name);
+    assert!(dump_path.exists(), "path: `{}`", dump_path.display());
+
+    opt.import_dump = Some(dump_path);
+    // NOTE: We shouldn't have to change the database path but I lost one hour
+    // because of a « bad path » error and that fixed it.
+    opt.db_path = temp.path().join("data.ms");
+
+    let mut server = Server::new_auth_with_options(opt, temp).await;
+    server.use_api_key("MASTER_KEY");
+
+    let (indexes, code) = server.list_indexes(None, None).await;
+    assert_eq!(code, 200, "{indexes}");
+
+    snapshot!(indexes["results"].as_array().unwrap().len(), @"1");
+    snapshot!(indexes["results"][0]["uid"], @r###""pets""###);
+    snapshot!(indexes["results"][0]["primaryKey"], @r###""id""###);
+
+    let (response, code) = server.get_features().await;
+    meili_snap::snapshot!(code, @"200 OK");
+    meili_snap::snapshot!(meili_snap::json_string!(response), @r###"
+    {
+      "vectorStore": true,
+      "metrics": false,
+      "logsRoute": false
+    }
+    "###);
+
+    let index = server.index("pets");
+
+    let (response, code) = index.settings().await;
+    meili_snap::snapshot!(code, @"200 OK");
+    meili_snap::snapshot!(meili_snap::json_string!(response), @r###"
+    {
+      "displayedAttributes": [
+        "*"
+      ],
+      "searchableAttributes": [
+        "*"
+      ],
+      "filterableAttributes": [],
+      "sortableAttributes": [],
+      "rankingRules": [
+        "words",
+        "typo",
+        "proximity",
+        "attribute",
+        "sort",
+        "exactness"
+      ],
+      "stopWords": [],
+      "nonSeparatorTokens": [],
+      "separatorTokens": [],
+      "dictionary": [],
+      "synonyms": {},
+      "distinctAttribute": null,
+      "proximityPrecision": "byWord",
+      "typoTolerance": {
+        "enabled": true,
+        "minWordSizeForTypos": {
+          "oneTypo": 5,
+          "twoTypos": 9
+        },
+        "disableOnWords": [],
+        "disableOnAttributes": []
+      },
+      "faceting": {
+        "maxValuesPerFacet": 100,
+        "sortFacetValuesBy": {
+          "*": "alpha"
+        }
+      },
+      "pagination": {
+        "maxTotalHits": 1000
+      },
+      "embedders": {
+        "doggo_embedder": {
+          "source": "huggingFace",
+          "model": "sentence-transformers/all-MiniLM-L6-v2",
+          "revision": "e4ce9877abf3edfe10b0d82785e83bdcb973e22e",
+          "documentTemplate": "{{doc.doggo}}"
+        }
+      },
+      "searchCutoffMs": null
+    }
+    "###);
+
+    index
+        .search(json!({"retrieveVectors": true}), |response, code| {
+            snapshot!(code, @"200 OK");
+            snapshot!(json_string!(response["hits"], { "[]._vectors.doggo_embedder.embeddings" => "[vector]" }), @r###"
+            [
+              {
+                "id": 0,
+                "doggo": "kefir",
+                "_vectors": {
+                  "doggo_embedder": {
+                    "embeddings": "[vector]",
+                    "regenerate": false
+                  }
+                }
+              },
+              {
+                "id": 1,
+                "doggo": "echo",
+                "_vectors": {
+                  "doggo_embedder": {
+                    "embeddings": "[vector]",
+                    "regenerate": false
+                  }
+                }
+              },
+              {
+                "id": 2,
+                "doggo": "intel",
+                "_vectors": {
+                  "doggo_embedder": {
+                    "embeddings": "[vector]",
+                    "regenerate": true
+                  }
+                }
+              },
+              {
+                "id": 3,
+                "doggo": "bill",
+                "_vectors": {
+                  "doggo_embedder": {
+                    "embeddings": "[vector]",
+                    "regenerate": true
+                  }
+                }
+              },
+              {
+                "id": 4,
+                "doggo": "max",
+                "_vectors": {
+                  "doggo_embedder": {
+                    "embeddings": "[vector]",
+                    "regenerate": true
+                  }
+                }
+              }
+            ]
+            "###);
+        })
+        .await;
+}

meilisearch/tests/dumps/snapshots/mod.rs/generate_and_import_dump_containing_vectors/1.snap

@@ -0,0 +1,25 @@
+---
+source: meilisearch/tests/dumps/mod.rs
+---
+{
+  "uid": 0,
+  "indexUid": "pets",
+  "status": "succeeded",
+  "type": "settingsUpdate",
+  "canceledBy": null,
+  "details": {
+    "embedders": {
+      "doggo_embedder": {
+        "source": "huggingFace",
+        "model": "sentence-transformers/all-MiniLM-L6-v2",
+        "revision": "e4ce9877abf3edfe10b0d82785e83bdcb973e22e",
+        "documentTemplate": "{{doc.doggo}}"
+      }
+    }
+  },
+  "error": null,
+  "duration": "[duration]",
+  "enqueuedAt": "[date]",
+  "startedAt": "[date]",
+  "finishedAt": "[date]"
+}

meilisearch/tests/dumps/snapshots/mod.rs/generate_and_import_dump_containing_vectors/2.snap

@@ -0,0 +1,19 @@
+---
+source: meilisearch/tests/dumps/mod.rs
+---
+{
+  "uid": 1,
+  "indexUid": "pets",
+  "status": "succeeded",
+  "type": "documentAdditionOrUpdate",
+  "canceledBy": null,
+  "details": {
+    "receivedDocuments": 5,
+    "indexedDocuments": 5
+  },
+  "error": null,
+  "duration": "[duration]",
+  "enqueuedAt": "[date]",
+  "startedAt": "[date]",
+  "finishedAt": "[date]"
+}

meilisearch/tests/integration.rs

@@ -13,6 +13,7 @@ mod snapshot;
 mod stats;
 mod swap_indexes;
 mod tasks;
+mod vector;
 
 // Tests are isolated by features in different modules to allow better readability, test
 // targetability, and improved incremental compilation times.

meilisearch/tests/search/errors.rs

@@ -167,6 +167,74 @@ async fn search_bad_hits_per_page() {
     "###);
 }
 
+#[actix_rt::test]
+async fn search_bad_attributes_to_retrieve() {
+    let server = Server::new().await;
+    let index = server.index("test");
+
+    let (response, code) = index.search_post(json!({"attributesToRetrieve": "doggo"})).await;
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Invalid value type at `.attributesToRetrieve`: expected an array, but found a string: `\"doggo\"`",
+      "code": "invalid_search_attributes_to_retrieve",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_search_attributes_to_retrieve"
+    }
+    "###);
+    // Can't make the `attributes_to_retrieve` fail with a get search since it'll accept anything as an array of strings.
+}
+
+#[actix_rt::test]
+async fn search_bad_retrieve_vectors() {
+    let server = Server::new().await;
+    let index = server.index("test");
+
+    let (response, code) = index.search_post(json!({"retrieveVectors": "doggo"})).await;
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Invalid value type at `.retrieveVectors`: expected a boolean, but found a string: `\"doggo\"`",
+      "code": "invalid_search_retrieve_vectors",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_search_retrieve_vectors"
+    }
+    "###);
+
+    let (response, code) = index.search_post(json!({"retrieveVectors": [true]})).await;
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Invalid value type at `.retrieveVectors`: expected a boolean, but found an array: `[true]`",
+      "code": "invalid_search_retrieve_vectors",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_search_retrieve_vectors"
+    }
+    "###);
+
+    let (response, code) = index.search_get("retrieveVectors=").await;
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Invalid value in parameter `retrieveVectors`: could not parse `` as a boolean, expected either `true` or `false`",
+      "code": "invalid_search_retrieve_vectors",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_search_retrieve_vectors"
+    }
+    "###);
+
+    let (response, code) = index.search_get("retrieveVectors=doggo").await;
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Invalid value in parameter `retrieveVectors`: could not parse `doggo` as a boolean, expected either `true` or `false`",
+      "code": "invalid_search_retrieve_vectors",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_search_retrieve_vectors"
+    }
+    "###);
+}
+
 #[actix_rt::test]
 async fn search_bad_attributes_to_crop() {
     let server = Server::new().await;

meilisearch/tests/search/hybrid.rs

@@ -124,29 +124,29 @@ async fn simple_search() {
 
     let (response, code) = index
         .search_post(
-            json!({"q": "Captain", "vector": [1.0, 1.0], "hybrid": {"semanticRatio": 0.2}}),
+            json!({"q": "Captain", "vector": [1.0, 1.0], "hybrid": {"semanticRatio": 0.2}, "retrieveVectors": true}),
         )
         .await;
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"], @r###"[{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]}},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]}},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]}}]"###);
+    snapshot!(response["hits"], @r###"[{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":{"embeddings":[[1.0,2.0]],"regenerate":false}}},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":{"embeddings":[[2.0,3.0]],"regenerate":false}}},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":{"embeddings":[[1.0,3.0]],"regenerate":false}}}]"###);
     snapshot!(response["semanticHitCount"], @"0");
 
     let (response, code) = index
         .search_post(
-            json!({"q": "Captain", "vector": [1.0, 1.0], "hybrid": {"semanticRatio": 0.5}, "showRankingScore": true}),
+            json!({"q": "Captain", "vector": [1.0, 1.0], "hybrid": {"semanticRatio": 0.5}, "showRankingScore": true, "retrieveVectors": true}),
         )
         .await;
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":0.990290343761444},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":0.9848484848484848},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":0.9472135901451112}]"###);
+    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":{"embeddings":[[2.0,3.0]],"regenerate":false}},"_rankingScore":0.990290343761444},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":{"embeddings":[[1.0,2.0]],"regenerate":false}},"_rankingScore":0.9848484848484848},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":{"embeddings":[[1.0,3.0]],"regenerate":false}},"_rankingScore":0.9472135901451112}]"###);
     snapshot!(response["semanticHitCount"], @"2");
 
     let (response, code) = index
         .search_post(
-            json!({"q": "Captain", "vector": [1.0, 1.0], "hybrid": {"semanticRatio": 0.8}, "showRankingScore": true}),
+            json!({"q": "Captain", "vector": [1.0, 1.0], "hybrid": {"semanticRatio": 0.8}, "showRankingScore": true, "retrieveVectors": true}),
         )
         .await;
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":0.990290343761444},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":0.974341630935669},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":0.9472135901451112}]"###);
+    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":{"embeddings":[[2.0,3.0]],"regenerate":false}},"_rankingScore":0.990290343761444},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":{"embeddings":[[1.0,2.0]],"regenerate":false}},"_rankingScore":0.974341630935669},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":{"embeddings":[[1.0,3.0]],"regenerate":false}},"_rankingScore":0.9472135901451112}]"###);
     snapshot!(response["semanticHitCount"], @"3");
 }
 
@@ -204,10 +204,10 @@ async fn distribution_shift() {
     let server = Server::new().await;
     let index = index_with_documents_user_provided(&server, &SIMPLE_SEARCH_DOCUMENTS_VEC).await;
 
-    let search = json!({"q": "Captain", "vector": [1.0, 1.0], "showRankingScore": true, "hybrid": {"semanticRatio": 1.0}});
+    let search = json!({"q": "Captain", "vector": [1.0, 1.0], "showRankingScore": true, "hybrid": {"semanticRatio": 1.0}, "retrieveVectors": true});
     let (response, code) = index.search_post(search.clone()).await;
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":0.990290343761444},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":0.974341630935669},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":0.9472135901451112}]"###);
+    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":{"embeddings":[[2.0,3.0]],"regenerate":false}},"_rankingScore":0.990290343761444},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":{"embeddings":[[1.0,2.0]],"regenerate":false}},"_rankingScore":0.974341630935669},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":{"embeddings":[[1.0,3.0]],"regenerate":false}},"_rankingScore":0.9472135901451112}]"###);
 
     let (response, code) = index
         .update_settings(json!({
@@ -228,7 +228,7 @@ async fn distribution_shift() {
 
     let (response, code) = index.search_post(search).await;
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":0.19161224365234375},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":1.1920928955078125e-7},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":1.1920928955078125e-7}]"###);
+    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":{"embeddings":[[2.0,3.0]],"regenerate":false}},"_rankingScore":0.19161224365234375},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":{"embeddings":[[1.0,2.0]],"regenerate":false}},"_rankingScore":1.1920928955078125e-7},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":{"embeddings":[[1.0,3.0]],"regenerate":false}},"_rankingScore":1.1920928955078125e-7}]"###);
 }
 
 #[actix_rt::test]
@@ -239,20 +239,23 @@ async fn highlighter() {
     let (response, code) = index
         .search_post(json!({"q": "Captain Marvel", "vector": [1.0, 1.0],
             "hybrid": {"semanticRatio": 0.2},
-            "attributesToHighlight": [
-                     "desc"
+           "retrieveVectors": true,
+           "attributesToHighlight": [
+                     "desc",
+                     "_vectors",
                    ],
-                   "highlightPreTag": "**BEGIN**",
-                   "highlightPostTag": "**END**"
+           "highlightPreTag": "**BEGIN**",
+           "highlightPostTag": "**END**",
         }))
         .await;
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_formatted":{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":["2.0","3.0"]}}},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_formatted":{"title":"Shazam!","desc":"a **BEGIN**Captain**END** **BEGIN**Marvel**END** ersatz","id":"1","_vectors":{"default":["1.0","3.0"]}}},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_formatted":{"title":"Captain Planet","desc":"He's not part of the **BEGIN**Marvel**END** Cinematic Universe","id":"2","_vectors":{"default":["1.0","2.0"]}}}]"###);
+    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":{"embeddings":[[2.0,3.0]],"regenerate":false}},"_formatted":{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3"}},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":{"embeddings":[[1.0,3.0]],"regenerate":false}},"_formatted":{"title":"Shazam!","desc":"a **BEGIN**Captain**END** **BEGIN**Marvel**END** ersatz","id":"1"}},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":{"embeddings":[[1.0,2.0]],"regenerate":false}},"_formatted":{"title":"Captain Planet","desc":"He's not part of the **BEGIN**Marvel**END** Cinematic Universe","id":"2"}}]"###);
     snapshot!(response["semanticHitCount"], @"0");
 
     let (response, code) = index
         .search_post(json!({"q": "Captain Marvel", "vector": [1.0, 1.0],
             "hybrid": {"semanticRatio": 0.8},
+            "retrieveVectors": true,
             "showRankingScore": true,
             "attributesToHighlight": [
                      "desc"
@@ -262,13 +265,14 @@ async fn highlighter() {
         }))
         .await;
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_formatted":{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":["2.0","3.0"]}},"_rankingScore":0.990290343761444},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_formatted":{"title":"Captain Planet","desc":"He's not part of the **BEGIN**Marvel**END** Cinematic Universe","id":"2","_vectors":{"default":["1.0","2.0"]}},"_rankingScore":0.974341630935669},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_formatted":{"title":"Shazam!","desc":"a **BEGIN**Captain**END** **BEGIN**Marvel**END** ersatz","id":"1","_vectors":{"default":["1.0","3.0"]}},"_rankingScore":0.9472135901451112}]"###);
+    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":{"embeddings":[[2.0,3.0]],"regenerate":false}},"_formatted":{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3"},"_rankingScore":0.990290343761444},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":{"embeddings":[[1.0,2.0]],"regenerate":false}},"_formatted":{"title":"Captain Planet","desc":"He's not part of the **BEGIN**Marvel**END** Cinematic Universe","id":"2"},"_rankingScore":0.974341630935669},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":{"embeddings":[[1.0,3.0]],"regenerate":false}},"_formatted":{"title":"Shazam!","desc":"a **BEGIN**Captain**END** **BEGIN**Marvel**END** ersatz","id":"1"},"_rankingScore":0.9472135901451112}]"###);
     snapshot!(response["semanticHitCount"], @"3");
 
     // no highlighting on full semantic
     let (response, code) = index
         .search_post(json!({"q": "Captain Marvel", "vector": [1.0, 1.0],
             "hybrid": {"semanticRatio": 1.0},
+            "retrieveVectors": true,
             "showRankingScore": true,
             "attributesToHighlight": [
                      "desc"
@@ -278,7 +282,7 @@ async fn highlighter() {
         }))
         .await;
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_formatted":{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":["2.0","3.0"]}},"_rankingScore":0.990290343761444},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_formatted":{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":["1.0","2.0"]}},"_rankingScore":0.974341630935669},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_formatted":{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":["1.0","3.0"]}},"_rankingScore":0.9472135901451112}]"###);
+    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":{"embeddings":[[2.0,3.0]],"regenerate":false}},"_formatted":{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3"},"_rankingScore":0.990290343761444},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":{"embeddings":[[1.0,2.0]],"regenerate":false}},"_formatted":{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2"},"_rankingScore":0.974341630935669},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":{"embeddings":[[1.0,3.0]],"regenerate":false}},"_formatted":{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1"},"_rankingScore":0.9472135901451112}]"###);
     snapshot!(response["semanticHitCount"], @"3");
 }
 
@@ -361,12 +365,12 @@ async fn single_document() {
 
     let (response, code) = index
     .search_post(
-        json!({"vector": [1.0, 3.0], "hybrid": {"semanticRatio": 1.0}, "showRankingScore": true}),
+        json!({"vector": [1.0, 3.0], "hybrid": {"semanticRatio": 1.0}, "showRankingScore": true, "retrieveVectors": true}),
     )
     .await;
 
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"][0], @r###"{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":1.0}"###);
+    snapshot!(response["hits"][0], @r###"{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":{"embeddings":[[1.0,3.0]],"regenerate":false}},"_rankingScore":1.0}"###);
     snapshot!(response["semanticHitCount"], @"1");
 }
 
@@ -377,25 +381,25 @@ async fn query_combination() {
 
     // search without query and vector, but with hybrid => still placeholder
     let (response, code) = index
-        .search_post(json!({"hybrid": {"semanticRatio": 1.0}, "showRankingScore": true}))
+        .search_post(json!({"hybrid": {"semanticRatio": 1.0}, "showRankingScore": true, "retrieveVectors": true}))
         .await;
 
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"], @r###"[{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":1.0},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":1.0},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":1.0}]"###);
+    snapshot!(response["hits"], @r###"[{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":{"embeddings":[[1.0,3.0]],"regenerate":false}},"_rankingScore":1.0},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":{"embeddings":[[1.0,2.0]],"regenerate":false}},"_rankingScore":1.0},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":{"embeddings":[[2.0,3.0]],"regenerate":false}},"_rankingScore":1.0}]"###);
     snapshot!(response["semanticHitCount"], @"null");
 
     // same with a different semantic ratio
     let (response, code) = index
-        .search_post(json!({"hybrid": {"semanticRatio": 0.76}, "showRankingScore": true}))
+        .search_post(json!({"hybrid": {"semanticRatio": 0.76}, "showRankingScore": true, "retrieveVectors": true}))
         .await;
 
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"], @r###"[{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":1.0},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":1.0},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":1.0}]"###);
+    snapshot!(response["hits"], @r###"[{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":{"embeddings":[[1.0,3.0]],"regenerate":false}},"_rankingScore":1.0},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":{"embeddings":[[1.0,2.0]],"regenerate":false}},"_rankingScore":1.0},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":{"embeddings":[[2.0,3.0]],"regenerate":false}},"_rankingScore":1.0}]"###);
     snapshot!(response["semanticHitCount"], @"null");
 
     // wrong vector dimensions
     let (response, code) = index
-    .search_post(json!({"vector": [1.0, 0.0, 1.0], "hybrid": {"semanticRatio": 1.0}, "showRankingScore": true}))
+    .search_post(json!({"vector": [1.0, 0.0, 1.0], "hybrid": {"semanticRatio": 1.0}, "showRankingScore": true, "retrieveVectors": true}))
     .await;
 
     snapshot!(code, @"400 Bad Request");
@@ -410,34 +414,34 @@ async fn query_combination() {
 
     // full vector
     let (response, code) = index
-    .search_post(json!({"vector": [1.0, 0.0], "hybrid": {"semanticRatio": 1.0}, "showRankingScore": true}))
+    .search_post(json!({"vector": [1.0, 0.0], "hybrid": {"semanticRatio": 1.0}, "showRankingScore": true, "retrieveVectors": true}))
     .await;
 
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":0.7773500680923462},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":0.7236068248748779},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":0.6581138968467712}]"###);
+    snapshot!(response["hits"], @r###"[{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":{"embeddings":[[2.0,3.0]],"regenerate":false}},"_rankingScore":0.7773500680923462},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":{"embeddings":[[1.0,2.0]],"regenerate":false}},"_rankingScore":0.7236068248748779},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":{"embeddings":[[1.0,3.0]],"regenerate":false}},"_rankingScore":0.6581138968467712}]"###);
     snapshot!(response["semanticHitCount"], @"3");
 
     // full keyword, without a query
     let (response, code) = index
-    .search_post(json!({"vector": [1.0, 0.0], "hybrid": {"semanticRatio": 0.0}, "showRankingScore": true}))
+    .search_post(json!({"vector": [1.0, 0.0], "hybrid": {"semanticRatio": 0.0}, "showRankingScore": true, "retrieveVectors": true}))
     .await;
 
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"], @r###"[{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":1.0},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":1.0},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":1.0}]"###);
+    snapshot!(response["hits"], @r###"[{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":{"embeddings":[[1.0,3.0]],"regenerate":false}},"_rankingScore":1.0},{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":{"embeddings":[[1.0,2.0]],"regenerate":false}},"_rankingScore":1.0},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":{"embeddings":[[2.0,3.0]],"regenerate":false}},"_rankingScore":1.0}]"###);
     snapshot!(response["semanticHitCount"], @"null");
 
     // query + vector, full keyword => keyword
     let (response, code) = index
-    .search_post(json!({"q": "Captain", "vector": [1.0, 0.0], "hybrid": {"semanticRatio": 0.0}, "showRankingScore": true}))
+    .search_post(json!({"q": "Captain", "vector": [1.0, 0.0], "hybrid": {"semanticRatio": 0.0}, "showRankingScore": true, "retrieveVectors": true}))
     .await;
 
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"], @r###"[{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":0.9848484848484848},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":[2.0,3.0]},"_rankingScore":0.9848484848484848},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":[1.0,3.0]},"_rankingScore":0.9242424242424242}]"###);
+    snapshot!(response["hits"], @r###"[{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":{"embeddings":[[1.0,2.0]],"regenerate":false}},"_rankingScore":0.9848484848484848},{"title":"Captain Marvel","desc":"a Shazam ersatz","id":"3","_vectors":{"default":{"embeddings":[[2.0,3.0]],"regenerate":false}},"_rankingScore":0.9848484848484848},{"title":"Shazam!","desc":"a Captain Marvel ersatz","id":"1","_vectors":{"default":{"embeddings":[[1.0,3.0]],"regenerate":false}},"_rankingScore":0.9242424242424242}]"###);
     snapshot!(response["semanticHitCount"], @"null");
 
     // query + vector, no hybrid keyword =>
     let (response, code) = index
-        .search_post(json!({"q": "Captain", "vector": [1.0, 0.0], "showRankingScore": true}))
+        .search_post(json!({"q": "Captain", "vector": [1.0, 0.0], "showRankingScore": true, "retrieveVectors": true}))
         .await;
 
     snapshot!(code, @"400 Bad Request");
@@ -453,7 +457,7 @@ async fn query_combination() {
     // full vector, without a vector => error
     let (response, code) = index
         .search_post(
-            json!({"q": "Captain", "hybrid": {"semanticRatio": 1.0}, "showRankingScore": true}),
+            json!({"q": "Captain", "hybrid": {"semanticRatio": 1.0}, "showRankingScore": true, "retrieveVectors": true}),
         )
         .await;
 
@@ -470,11 +474,93 @@ async fn query_combination() {
     // hybrid without a vector => full keyword
     let (response, code) = index
         .search_post(
-            json!({"q": "Planet", "hybrid": {"semanticRatio": 0.99}, "showRankingScore": true}),
+            json!({"q": "Planet", "hybrid": {"semanticRatio": 0.99}, "showRankingScore": true, "retrieveVectors": true}),
         )
         .await;
 
     snapshot!(code, @"200 OK");
-    snapshot!(response["hits"], @r###"[{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":[1.0,2.0]},"_rankingScore":0.9242424242424242}]"###);
+    snapshot!(response["hits"], @r###"[{"title":"Captain Planet","desc":"He's not part of the Marvel Cinematic Universe","id":"2","_vectors":{"default":{"embeddings":[[1.0,2.0]],"regenerate":false}},"_rankingScore":0.9242424242424242}]"###);
     snapshot!(response["semanticHitCount"], @"0");
 }
+
+#[actix_rt::test]
+async fn retrieve_vectors() {
+    let server = Server::new().await;
+    let index = index_with_documents_hf(&server, &SIMPLE_SEARCH_DOCUMENTS).await;
+
+    let (response, code) = index
+        .search_post(
+            json!({"q": "Captain", "hybrid": {"semanticRatio": 0.2}, "retrieveVectors": true}),
+        )
+        .await;
+    snapshot!(code, @"200 OK");
+    insta::assert_json_snapshot!(response["hits"], {"[]._vectors.default.embeddings" => "[vectors]"},  @r###"
+    [
+      {
+        "title": "Captain Planet",
+        "desc": "He's not part of the Marvel Cinematic Universe",
+        "id": "2",
+        "_vectors": {
+          "default": {
+            "embeddings": "[vectors]",
+            "regenerate": true
+          }
+        }
+      },
+      {
+        "title": "Captain Marvel",
+        "desc": "a Shazam ersatz",
+        "id": "3",
+        "_vectors": {
+          "default": {
+            "embeddings": "[vectors]",
+            "regenerate": true
+          }
+        }
+      },
+      {
+        "title": "Shazam!",
+        "desc": "a Captain Marvel ersatz",
+        "id": "1",
+        "_vectors": {
+          "default": {
+            "embeddings": "[vectors]",
+            "regenerate": true
+          }
+        }
+      }
+    ]
+    "###);
+
+    // remove `_vectors` from displayed attributes
+    let (response, code) =
+        index.update_settings(json!({ "displayedAttributes": ["id", "title", "desc"]} )).await;
+    assert_eq!(202, code, "{:?}", response);
+    index.wait_task(response.uid()).await;
+
+    let (response, code) = index
+        .search_post(
+            json!({"q": "Captain", "hybrid": {"semanticRatio": 0.2}, "retrieveVectors": true}),
+        )
+        .await;
+    snapshot!(code, @"200 OK");
+    insta::assert_json_snapshot!(response["hits"], {"[]._vectors.default.embeddings" => "[vectors]"},  @r###"
+    [
+      {
+        "title": "Captain Planet",
+        "desc": "He's not part of the Marvel Cinematic Universe",
+        "id": "2"
+      },
+      {
+        "title": "Captain Marvel",
+        "desc": "a Shazam ersatz",
+        "id": "3"
+      },
+      {
+        "title": "Shazam!",
+        "desc": "a Captain Marvel ersatz",
+        "id": "1"
+      }
+    ]
+    "###);
+}

meilisearch/tests/search/mod.rs

@@ -1290,21 +1290,38 @@ async fn experimental_feature_vector_store() {
     index.add_documents(json!(documents), None).await;
     index.wait_task(0).await;
 
-    let (response, code) = index
-        .search_post(json!({
+    index
+        .search(json!({
             "vector": [1.0, 2.0, 3.0],
             "showRankingScore": true
-        }))
+        }), |response, code|{
+            meili_snap::snapshot!(code, @"400 Bad Request");
+            meili_snap::snapshot!(meili_snap::json_string!(response), @r###"
+            {
+              "message": "Passing `vector` as a parameter requires enabling the `vector store` experimental feature. See https://github.com/meilisearch/product/discussions/677",
+              "code": "feature_not_enabled",
+              "type": "invalid_request",
+              "link": "https://docs.meilisearch.com/errors#feature_not_enabled"
+            }
+            "###);
+        })
+        .await;
+    index
+        .search(json!({
+            "retrieveVectors": true,
+            "showRankingScore": true
+        }), |response, code|{
+            meili_snap::snapshot!(code, @"400 Bad Request");
+            meili_snap::snapshot!(meili_snap::json_string!(response), @r###"
+            {
+              "message": "Passing `retrieveVectors` as a parameter requires enabling the `vector store` experimental feature. See https://github.com/meilisearch/product/discussions/677",
+              "code": "feature_not_enabled",
+              "type": "invalid_request",
+              "link": "https://docs.meilisearch.com/errors#feature_not_enabled"
+            }
+            "###);
+        })
         .await;
-    meili_snap::snapshot!(code, @"400 Bad Request");
-    meili_snap::snapshot!(meili_snap::json_string!(response), @r###"
-    {
-      "message": "Passing `vector` as a query parameter requires enabling the `vector store` experimental feature. See https://github.com/meilisearch/product/discussions/677",
-      "code": "feature_not_enabled",
-      "type": "invalid_request",
-      "link": "https://docs.meilisearch.com/errors#feature_not_enabled"
-    }
-    "###);
 
     let (response, code) = server.set_features(json!({"vectorStore": true})).await;
     meili_snap::snapshot!(code, @"200 OK");
@@ -1337,6 +1354,7 @@ async fn experimental_feature_vector_store() {
         .search_post(json!({
             "vector": [1.0, 2.0, 3.0],
             "showRankingScore": true,
+            "retrieveVectors": true,
         }))
         .await;
 
@@ -1348,11 +1366,16 @@ async fn experimental_feature_vector_store() {
         "title": "Shazam!",
         "id": "287947",
         "_vectors": {
-          "manual": [
-            1.0,
-            2.0,
-            3.0
-          ]
+          "manual": {
+            "embeddings": [
+              [
+                1.0,
+                2.0,
+                3.0
+              ]
+            ],
+            "regenerate": false
+          }
         },
         "_rankingScore": 1.0
       },
@@ -1360,11 +1383,16 @@ async fn experimental_feature_vector_store() {
         "title": "Captain Marvel",
         "id": "299537",
         "_vectors": {
-          "manual": [
-            1.0,
-            2.0,
-            54.0
-          ]
+          "manual": {
+            "embeddings": [
+              [
+                1.0,
+                2.0,
+                54.0
+              ]
+            ],
+            "regenerate": false
+          }
         },
         "_rankingScore": 0.9129111766815186
       },
@@ -1372,11 +1400,16 @@ async fn experimental_feature_vector_store() {
         "title": "Gläss",
         "id": "450465",
         "_vectors": {
-          "manual": [
-            -100.0,
-            340.0,
-            90.0
-          ]
+          "manual": {
+            "embeddings": [
+              [
+                -100.0,
+                340.0,
+                90.0
+              ]
+            ],
+            "regenerate": false
+          }
         },
         "_rankingScore": 0.8106412887573242
       },
@@ -1384,11 +1417,16 @@ async fn experimental_feature_vector_store() {
         "title": "How to Train Your Dragon: The Hidden World",
         "id": "166428",
         "_vectors": {
-          "manual": [
-            -100.0,
-            231.0,
-            32.0
-          ]
+          "manual": {
+            "embeddings": [
+              [
+                -100.0,
+                231.0,
+                32.0
+              ]
+            ],
+            "regenerate": false
+          }
         },
         "_rankingScore": 0.7412010431289673
       },
@@ -1396,11 +1434,16 @@ async fn experimental_feature_vector_store() {
         "title": "Escape Room",
         "id": "522681",
         "_vectors": {
-          "manual": [
-            10.0,
-            -23.0,
-            32.0
-          ]
+          "manual": {
+            "embeddings": [
+              [
+                10.0,
+                -23.0,
+                32.0
+              ]
+            ],
+            "regenerate": false
+          }
         },
         "_rankingScore": 0.6972063183784485
       }

meilisearch/tests/similar/errors.rs

@@ -756,3 +756,54 @@ async fn filter_reserved_geo_point_string() {
         })
         .await;
 }
+
+#[actix_rt::test]
+async fn similar_bad_retrieve_vectors() {
+    let server = Server::new().await;
+    server.set_features(json!({"vectorStore": true})).await;
+    let index = server.index("test");
+
+    let (response, code) = index.similar_post(json!({"retrieveVectors": "doggo"})).await;
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Invalid value type at `.retrieveVectors`: expected a boolean, but found a string: `\"doggo\"`",
+      "code": "invalid_similar_retrieve_vectors",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_similar_retrieve_vectors"
+    }
+    "###);
+
+    let (response, code) = index.similar_post(json!({"retrieveVectors": [true]})).await;
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Invalid value type at `.retrieveVectors`: expected a boolean, but found an array: `[true]`",
+      "code": "invalid_similar_retrieve_vectors",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_similar_retrieve_vectors"
+    }
+    "###);
+
+    let (response, code) = index.similar_get("retrieveVectors=").await;
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Invalid value in parameter `retrieveVectors`: could not parse `` as a boolean, expected either `true` or `false`",
+      "code": "invalid_similar_retrieve_vectors",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_similar_retrieve_vectors"
+    }
+    "###);
+
+    let (response, code) = index.similar_get("retrieveVectors=doggo").await;
+    snapshot!(code, @"400 Bad Request");
+    snapshot!(json_string!(response), @r###"
+    {
+      "message": "Invalid value in parameter `retrieveVectors`: could not parse `doggo` as a boolean, expected either `true` or `false`",
+      "code": "invalid_similar_retrieve_vectors",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_similar_retrieve_vectors"
+    }
+    "###);
+}

meilisearch/tests/similar/mod.rs

@@ -78,7 +78,7 @@ async fn basic() {
     index.wait_task(value.uid()).await;
 
     index
-        .similar(json!({"id": 143}), |response, code| {
+        .similar(json!({"id": 143, "retrieveVectors": true}), |response, code| {
             snapshot!(code, @"200 OK");
             snapshot!(json_string!(response["hits"]), @r###"
             [
@@ -87,11 +87,16 @@ async fn basic() {
                 "release_year": 2019,
                 "id": "522681",
                 "_vectors": {
-                  "manual": [
-                    0.1,
-                    0.6,
-                    0.8
-                  ]
+                  "manual": {
+                    "embeddings": [
+                      [
+                        0.10000000149011612,
+                        0.6000000238418579,
+                        0.800000011920929
+                      ]
+                    ],
+                    "regenerate": false
+                  }
                 }
               },
               {
@@ -99,11 +104,16 @@ async fn basic() {
                 "release_year": 2019,
                 "id": "299537",
                 "_vectors": {
-                  "manual": [
-                    0.6,
-                    0.8,
-                    -0.2
-                  ]
+                  "manual": {
+                    "embeddings": [
+                      [
+                        0.6000000238418579,
+                        0.800000011920929,
+                        -0.20000000298023224
+                      ]
+                    ],
+                    "regenerate": false
+                  }
                 }
               },
               {
@@ -111,11 +121,16 @@ async fn basic() {
                 "release_year": 2019,
                 "id": "166428",
                 "_vectors": {
-                  "manual": [
-                    0.7,
-                    0.7,
-                    -0.4
-                  ]
+                  "manual": {
+                    "embeddings": [
+                      [
+                        0.699999988079071,
+                        0.699999988079071,
+                        -0.4000000059604645
+                      ]
+                    ],
+                    "regenerate": false
+                  }
                 }
               },
               {
@@ -123,11 +138,16 @@ async fn basic() {
                 "release_year": 2019,
                 "id": "287947",
                 "_vectors": {
-                  "manual": [
-                    0.8,
-                    0.4,
-                    -0.5
-                  ]
+                  "manual": {
+                    "embeddings": [
+                      [
+                        0.800000011920929,
+                        0.4000000059604645,
+                        -0.5
+                      ]
+                    ],
+                    "regenerate": false
+                  }
                 }
               }
             ]
@@ -136,7 +156,7 @@ async fn basic() {
         .await;
 
     index
-        .similar(json!({"id": "299537"}), |response, code| {
+        .similar(json!({"id": "299537", "retrieveVectors": true}), |response, code| {
             snapshot!(code, @"200 OK");
             snapshot!(json_string!(response["hits"]), @r###"
             [
@@ -145,11 +165,16 @@ async fn basic() {
                 "release_year": 2019,
                 "id": "166428",
                 "_vectors": {
-                  "manual": [
-                    0.7,
-                    0.7,
-                    -0.4
-                  ]
+                  "manual": {
+                    "embeddings": [
+                      [
+                        0.699999988079071,
+                        0.699999988079071,
+                        -0.4000000059604645
+                      ]
+                    ],
+                    "regenerate": false
+                  }
                 }
               },
               {
@@ -157,11 +182,16 @@ async fn basic() {
                 "release_year": 2019,
                 "id": "287947",
                 "_vectors": {
-                  "manual": [
-                    0.8,
-                    0.4,
-                    -0.5
-                  ]
+                  "manual": {
+                    "embeddings": [
+                      [
+                        0.800000011920929,
+                        0.4000000059604645,
+                        -0.5
+                      ]
+                    ],
+                    "regenerate": false
+                  }
                 }
               },
               {
@@ -169,11 +199,16 @@ async fn basic() {
                 "release_year": 2019,
                 "id": "522681",
                 "_vectors": {
-                  "manual": [
-                    0.1,
-                    0.6,
-                    0.8
-                  ]
+                  "manual": {
+                    "embeddings": [
+                      [
+                        0.10000000149011612,
+                        0.6000000238418579,
+                        0.800000011920929
+                      ]
+                    ],
+                    "regenerate": false
+                  }
                 }
               },
               {
@@ -181,11 +216,16 @@ async fn basic() {
                 "release_year": 1930,
                 "id": "143",
                 "_vectors": {
-                  "manual": [
-                    -0.5,
-                    0.3,
-                    0.85
-                  ]
+                  "manual": {
+                    "embeddings": [
+                      [
+                        -0.5,
+                        0.30000001192092896,
+                        0.8500000238418579
+                      ]
+                    ],
+                    "regenerate": false
+                  }
                 }
               }
             ]
@@ -228,7 +268,7 @@ async fn ranking_score_threshold() {
 
     index
         .similar(
-            json!({"id": 143, "showRankingScore": true, "rankingScoreThreshold": 0}),
+            json!({"id": 143, "showRankingScore": true, "rankingScoreThreshold": 0, "retrieveVectors": true}),
             |response, code| {
                 snapshot!(code, @"200 OK");
                 meili_snap::snapshot!(meili_snap::json_string!(response["estimatedTotalHits"]), @"4");
@@ -239,11 +279,16 @@ async fn ranking_score_threshold() {
                     "release_year": 2019,
                     "id": "522681",
                     "_vectors": {
-                      "manual": [
-                        0.1,
-                        0.6,
-                        0.8
-                      ]
+                      "manual": {
+                        "embeddings": [
+                          [
+                            0.10000000149011612,
+                            0.6000000238418579,
+                            0.800000011920929
+                          ]
+                        ],
+                        "regenerate": false
+                      }
                     },
                     "_rankingScore": 0.890957772731781
                   },
@@ -252,11 +297,16 @@ async fn ranking_score_threshold() {
                     "release_year": 2019,
                     "id": "299537",
                     "_vectors": {
-                      "manual": [
-                        0.6,
-                        0.8,
-                        -0.2
-                      ]
+                      "manual": {
+                        "embeddings": [
+                          [
+                            0.6000000238418579,
+                            0.800000011920929,
+                            -0.20000000298023224
+                          ]
+                        ],
+                        "regenerate": false
+                      }
                     },
                     "_rankingScore": 0.39060014486312866
                   },
@@ -265,11 +315,16 @@ async fn ranking_score_threshold() {
                     "release_year": 2019,
                     "id": "166428",
                     "_vectors": {
-                      "manual": [
-                        0.7,
-                        0.7,
-                        -0.4
-                      ]
+                      "manual": {
+                        "embeddings": [
+                          [
+                            0.699999988079071,
+                            0.699999988079071,
+                            -0.4000000059604645
+                          ]
+                        ],
+                        "regenerate": false
+                      }
                     },
                     "_rankingScore": 0.2819308042526245
                   },
@@ -278,11 +333,16 @@ async fn ranking_score_threshold() {
                     "release_year": 2019,
                     "id": "287947",
                     "_vectors": {
-                      "manual": [
-                        0.8,
-                        0.4,
-                        -0.5
-                      ]
+                      "manual": {
+                        "embeddings": [
+                          [
+                            0.800000011920929,
+                            0.4000000059604645,
+                            -0.5
+                          ]
+                        ],
+                        "regenerate": false
+                      }
                     },
                     "_rankingScore": 0.1662663221359253
                   }
@@ -294,7 +354,7 @@ async fn ranking_score_threshold() {
 
     index
         .similar(
-            json!({"id": 143, "showRankingScore": true, "rankingScoreThreshold": 0.2}),
+            json!({"id": 143, "showRankingScore": true, "rankingScoreThreshold": 0.2, "retrieveVectors": true}),
             |response, code| {
                 snapshot!(code, @"200 OK");
                 meili_snap::snapshot!(meili_snap::json_string!(response["estimatedTotalHits"]), @"3");
@@ -305,11 +365,16 @@ async fn ranking_score_threshold() {
                     "release_year": 2019,
                     "id": "522681",
                     "_vectors": {
-                      "manual": [
-                        0.1,
-                        0.6,
-                        0.8
-                      ]
+                      "manual": {
+                        "embeddings": [
+                          [
+                            0.10000000149011612,
+                            0.6000000238418579,
+                            0.800000011920929
+                          ]
+                        ],
+                        "regenerate": false
+                      }
                     },
                     "_rankingScore": 0.890957772731781
                   },
@@ -318,11 +383,16 @@ async fn ranking_score_threshold() {
                     "release_year": 2019,
                     "id": "299537",
                     "_vectors": {
-                      "manual": [
-                        0.6,
-                        0.8,
-                        -0.2
-                      ]
+                      "manual": {
+                        "embeddings": [
+                          [
+                            0.6000000238418579,
+                            0.800000011920929,
+                            -0.20000000298023224
+                          ]
+                        ],
+                        "regenerate": false
+                      }
                     },
                     "_rankingScore": 0.39060014486312866
                   },
@@ -331,11 +401,16 @@ async fn ranking_score_threshold() {
                     "release_year": 2019,
                     "id": "166428",
                     "_vectors": {
-                      "manual": [
-                        0.7,
-                        0.7,
-                        -0.4
-                      ]
+                      "manual": {
+                        "embeddings": [
+                          [
+                            0.699999988079071,
+                            0.699999988079071,
+                            -0.4000000059604645
+                          ]
+                        ],
+                        "regenerate": false
+                      }
                     },
                     "_rankingScore": 0.2819308042526245
                   }
@@ -347,7 +422,7 @@ async fn ranking_score_threshold() {
 
     index
         .similar(
-            json!({"id": 143, "showRankingScore": true, "rankingScoreThreshold": 0.3}),
+            json!({"id": 143, "showRankingScore": true, "rankingScoreThreshold": 0.3, "retrieveVectors": true}),
             |response, code| {
                 snapshot!(code, @"200 OK");
                 meili_snap::snapshot!(meili_snap::json_string!(response["estimatedTotalHits"]), @"2");
@@ -358,11 +433,16 @@ async fn ranking_score_threshold() {
                     "release_year": 2019,
                     "id": "522681",
                     "_vectors": {
-                      "manual": [
-                        0.1,
-                        0.6,
-                        0.8
-                      ]
+                      "manual": {
+                        "embeddings": [
+                          [
+                            0.10000000149011612,
+                            0.6000000238418579,
+                            0.800000011920929
+                          ]
+                        ],
+                        "regenerate": false
+                      }
                     },
                     "_rankingScore": 0.890957772731781
                   },
@@ -371,11 +451,16 @@ async fn ranking_score_threshold() {
                     "release_year": 2019,
                     "id": "299537",
                     "_vectors": {
-                      "manual": [
-                        0.6,
-                        0.8,
-                        -0.2
-                      ]
+                      "manual": {
+                        "embeddings": [
+                          [
+                            0.6000000238418579,
+                            0.800000011920929,
+                            -0.20000000298023224
+                          ]
+                        ],
+                        "regenerate": false
+                      }
                     },
                     "_rankingScore": 0.39060014486312866
                   }
@@ -387,7 +472,7 @@ async fn ranking_score_threshold() {
 
     index
         .similar(
-            json!({"id": 143, "showRankingScore": true, "rankingScoreThreshold": 0.6}),
+            json!({"id": 143, "showRankingScore": true, "rankingScoreThreshold": 0.6, "retrieveVectors": true}),
             |response, code| {
                 snapshot!(code, @"200 OK");
                 meili_snap::snapshot!(meili_snap::json_string!(response["estimatedTotalHits"]), @"1");
@@ -398,11 +483,16 @@ async fn ranking_score_threshold() {
                     "release_year": 2019,
                     "id": "522681",
                     "_vectors": {
-                      "manual": [
-                        0.1,
-                        0.6,
-                        0.8
-                      ]
+                      "manual": {
+                        "embeddings": [
+                          [
+                            0.10000000149011612,
+                            0.6000000238418579,
+                            0.800000011920929
+                          ]
+                        ],
+                        "regenerate": false
+                      }
                     },
                     "_rankingScore": 0.890957772731781
                   }
@@ -414,7 +504,7 @@ async fn ranking_score_threshold() {
 
     index
         .similar(
-            json!({"id": 143, "showRankingScore": true, "rankingScoreThreshold": 0.9}),
+            json!({"id": 143, "showRankingScore": true, "rankingScoreThreshold": 0.9, "retrieveVectors": true}),
             |response, code| {
                 snapshot!(code, @"200 OK");
                 snapshot!(json_string!(response["hits"]), @"[]");
@@ -456,71 +546,97 @@ async fn filter() {
     index.wait_task(value.uid()).await;
 
     index
-        .similar(json!({"id": 522681, "filter": "release_year = 2019"}), |response, code| {
-            snapshot!(code, @"200 OK");
-            snapshot!(json_string!(response["hits"]), @r###"
-            [
-              {
-                "title": "Captain Marvel",
-                "release_year": 2019,
-                "id": "299537",
-                "_vectors": {
-                  "manual": [
-                    0.6,
-                    0.8,
-                    -0.2
-                  ]
-                }
-              },
-              {
-                "title": "How to Train Your Dragon: The Hidden World",
-                "release_year": 2019,
-                "id": "166428",
-                "_vectors": {
-                  "manual": [
-                    0.7,
-                    0.7,
-                    -0.4
-                  ]
-                }
-              },
-              {
-                "title": "Shazam!",
-                "release_year": 2019,
-                "id": "287947",
-                "_vectors": {
-                  "manual": [
-                    0.8,
-                    0.4,
-                    -0.5
-                  ]
-                }
-              }
-            ]
-            "###);
-        })
+        .similar(
+            json!({"id": 522681, "filter": "release_year = 2019", "retrieveVectors": true}),
+            |response, code| {
+                snapshot!(code, @"200 OK");
+                snapshot!(json_string!(response["hits"]), @r###"
+                [
+                  {
+                    "title": "Captain Marvel",
+                    "release_year": 2019,
+                    "id": "299537",
+                    "_vectors": {
+                      "manual": {
+                        "embeddings": [
+                          [
+                            0.6000000238418579,
+                            0.800000011920929,
+                            -0.20000000298023224
+                          ]
+                        ],
+                        "regenerate": false
+                      }
+                    }
+                  },
+                  {
+                    "title": "How to Train Your Dragon: The Hidden World",
+                    "release_year": 2019,
+                    "id": "166428",
+                    "_vectors": {
+                      "manual": {
+                        "embeddings": [
+                          [
+                            0.699999988079071,
+                            0.699999988079071,
+                            -0.4000000059604645
+                          ]
+                        ],
+                        "regenerate": false
+                      }
+                    }
+                  },
+                  {
+                    "title": "Shazam!",
+                    "release_year": 2019,
+                    "id": "287947",
+                    "_vectors": {
+                      "manual": {
+                        "embeddings": [
+                          [
+                            0.800000011920929,
+                            0.4000000059604645,
+                            -0.5
+                          ]
+                        ],
+                        "regenerate": false
+                      }
+                    }
+                  }
+                ]
+                "###);
+            },
+        )
         .await;
 
     index
-        .similar(json!({"id": 522681, "filter": "release_year < 2000"}), |response, code| {
-            snapshot!(code, @"200 OK");
-            snapshot!(json_string!(response["hits"]), @r###"
-            [
-              {
-                "title": "All Quiet on the Western Front",
-                "release_year": 1930,
-                "id": "143",
-                "_vectors": {
-                  "manual": [
-                    -0.5,
-                    0.3,
-                    0.85
-                  ]
-                }
-              }
-            ]
-            "###);
-        })
+        .similar(
+            json!({"id": 522681, "filter": "release_year < 2000", "retrieveVectors": true}),
+            |response, code| {
+                snapshot!(code, @"200 OK");
+                snapshot!(json_string!(response["hits"]), @r###"
+                [
+                  {
+                    "title": "All Quiet on the Western Front",
+                    "release_year": 1930,
+                    "id": "143",
+                    "_vectors": {
+                      "manual": {
+                        "embeddings": [
+                          [
+                            -0.5,
+                            0.30000001192092896,
+                            0.8500000238418579
+                          ]
+                        ],
+                        "regenerate": false
+                      }
+                    }
+                  }
+                ]
+                "###);
+            },
+        )
         .await;
 }
 
@@ -557,7 +673,7 @@ async fn limit_and_offset() {
     index.wait_task(value.uid()).await;
 
     index
-        .similar(json!({"id": 143, "limit": 1}), |response, code| {
+        .similar(json!({"id": 143, "limit": 1, "retrieveVectors": true}), |response, code| {
             snapshot!(code, @"200 OK");
             snapshot!(json_string!(response["hits"]), @r###"
             [
@@ -566,11 +682,16 @@ async fn limit_and_offset() {
                 "release_year": 2019,
                 "id": "522681",
                 "_vectors": {
-                  "manual": [
-                    0.1,
-                    0.6,
-                    0.8
-                  ]
+                  "manual": {
+                    "embeddings": [
+                      [
+                        0.10000000149011612,
+                        0.6000000238418579,
+                        0.800000011920929
+                      ]
+                    ],
+                    "regenerate": false
+                  }
                 }
               }
             ]
@@ -579,24 +700,32 @@ async fn limit_and_offset() {
         .await;
 
     index
-        .similar(json!({"id": 143, "limit": 1, "offset": 1}), |response, code| {
-            snapshot!(code, @"200 OK");
-            snapshot!(json_string!(response["hits"]), @r###"
-            [
-              {
-                "title": "Captain Marvel",
-                "release_year": 2019,
-                "id": "299537",
-                "_vectors": {
-                  "manual": [
-                    0.6,
-                    0.8,
-                    -0.2
-                  ]
-                }
-              }
-            ]
-            "###);
-        })
+        .similar(
+            json!({"id": 143, "limit": 1, "offset": 1, "retrieveVectors": true}),
+            |response, code| {
+                snapshot!(code, @"200 OK");
+                snapshot!(json_string!(response["hits"]), @r###"
+                [
+                  {
+                    "title": "Captain Marvel",
+                    "release_year": 2019,
+                    "id": "299537",
+                    "_vectors": {
+                      "manual": {
+                        "embeddings": [
+                          [
+                            0.6000000238418579,
+                            0.800000011920929,
+                            -0.20000000298023224
+                          ]
+                        ],
+                        "regenerate": false
+                      }
+                    }
+                  }
+                ]
+                "###);
+            },
+        )
         .await;
 }

meilisearch/tests/vector/mod.rs

@@ -0,0 +1,227 @@
+mod settings;
+
+use meili_snap::{json_string, snapshot};
+
+use crate::common::index::Index;
+use crate::common::{GetAllDocumentsOptions, Server};
+use crate::json;
+
+#[actix_rt::test]
+async fn add_remove_user_provided() {
+    let server = Server::new().await;
+    let index = server.index("doggo");
+    let (value, code) = server.set_features(json!({"vectorStore": true})).await;
+    snapshot!(code, @"200 OK");
+    snapshot!(value, @r###"
+    {
+      "vectorStore": true,
+      "metrics": false,
+      "logsRoute": false
+    }
+    "###);
+
+    let (response, code) = index
+        .update_settings(json!({
+          "embedders": {
+              "manual": {
+                  "source": "userProvided",
+                  "dimensions": 3,
+              }
+          },
+        }))
+        .await;
+    snapshot!(code, @"202 Accepted");
+    server.wait_task(response.uid()).await;
+
+    let documents = json!([
+      {"id": 0, "name": "kefir", "_vectors": { "manual": [0, 0, 0] }},
+      {"id": 1, "name": "echo", "_vectors": { "manual": [1, 1, 1] }},
+    ]);
+    let (value, code) = index.add_documents(documents, None).await;
+    snapshot!(code, @"202 Accepted");
+    index.wait_task(value.uid()).await;
+
+    let (documents, _code) = index
+        .get_all_documents(GetAllDocumentsOptions { retrieve_vectors: true, ..Default::default() })
+        .await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "results": [
+        {
+          "id": 0,
+          "name": "kefir",
+          "_vectors": {
+            "manual": {
+              "embeddings": [
+                [
+                  0.0,
+                  0.0,
+                  0.0
+                ]
+              ],
+              "regenerate": false
+            }
+          }
+        },
+        {
+          "id": 1,
+          "name": "echo",
+          "_vectors": {
+            "manual": {
+              "embeddings": [
+                [
+                  1.0,
+                  1.0,
+                  1.0
+                ]
+              ],
+              "regenerate": false
+            }
+          }
+        }
+      ],
+      "offset": 0,
+      "limit": 20,
+      "total": 2
+    }
+    "###);
+
+    let documents = json!([
+      {"id": 0, "name": "kefir", "_vectors": { "manual": [10, 10, 10] }},
+      {"id": 1, "name": "echo", "_vectors": { "manual": null }},
+    ]);
+    let (value, code) = index.add_documents(documents, None).await;
+    snapshot!(code, @"202 Accepted");
+    index.wait_task(value.uid()).await;
+
+    let (documents, _code) = index
+        .get_all_documents(GetAllDocumentsOptions { retrieve_vectors: true, ..Default::default() })
+        .await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "results": [
+        {
+          "id": 0,
+          "name": "kefir",
+          "_vectors": {
+            "manual": {
+              "embeddings": [
+                [
+                  10.0,
+                  10.0,
+                  10.0
+                ]
+              ],
+              "regenerate": false
+            }
+          }
+        },
+        {
+          "id": 1,
+          "name": "echo",
+          "_vectors": {}
+        }
+      ],
+      "offset": 0,
+      "limit": 20,
+      "total": 2
+    }
+    "###);
+
+    let (value, code) = index.delete_document(0).await;
+    snapshot!(code, @"202 Accepted");
+    index.wait_task(value.uid()).await;
+
+    let (documents, _code) = index
+        .get_all_documents(GetAllDocumentsOptions { retrieve_vectors: true, ..Default::default() })
+        .await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "results": [
+        {
+          "id": 1,
+          "name": "echo",
+          "_vectors": {}
+        }
+      ],
+      "offset": 0,
+      "limit": 20,
+      "total": 1
+    }
+    "###);
+}
+
+async fn generate_default_user_provided_documents(server: &Server) -> Index {
+    let index = server.index("doggo");
+    let (value, code) = server.set_features(json!({"vectorStore": true})).await;
+    snapshot!(code, @"200 OK");
+    snapshot!(value, @r###"
+    {
+      "vectorStore": true,
+      "metrics": false,
+      "logsRoute": false
+    }
+    "###);
+
+    let (response, code) = index
+        .update_settings(json!({
+          "embedders": {
+              "manual": {
+                  "source": "userProvided",
+                  "dimensions": 3,
+              }
+          },
+        }))
+        .await;
+    snapshot!(code, @"202 Accepted");
+    server.wait_task(response.uid()).await;
+
+    let documents = json!([
+      {"id": 0, "name": "kefir", "_vectors": { "manual": [0, 0, 0] }},
+      {"id": 1, "name": "echo", "_vectors": { "manual": [1, 1, 1] }},
+      {"id": 2, "name": "billou", "_vectors": { "manual": [[2, 2, 2], [2, 2, 3]] }},
+      {"id": 3, "name": "intel", "_vectors": { "manual": { "regenerate": false, "embeddings": [3, 3, 3] }}},
+      {"id": 4, "name": "max", "_vectors": { "manual": { "regenerate": false, "embeddings": [[4, 4, 4], [4, 4, 5]] }}},
+    ]);
+    let (value, code) = index.add_documents(documents, None).await;
+    snapshot!(code, @"202 Accepted");
+    index.wait_task(value.uid()).await;
+
+    index
+}
+
+#[actix_rt::test]
+async fn clear_documents() {
+    let server = Server::new().await;
+    let index = generate_default_user_provided_documents(&server).await;
+
+    let (value, _code) = index.clear_all_documents().await;
+    index.wait_task(value.uid()).await;
+
+    // Make sure the documents DB has been cleared
+    let (documents, _code) = index
+        .get_all_documents(GetAllDocumentsOptions { retrieve_vectors: true, ..Default::default() })
+        .await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "results": [],
+      "offset": 0,
+      "limit": 20,
+      "total": 0
+    }
+    "###);
+
+    // Make sure the arroy DB has been cleared
+    let (documents, _code) = index.search_post(json!({ "vector": [1, 1, 1] })).await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "hits": [],
+      "query": "",
+      "processingTimeMs": 0,
+      "limit": 20,
+      "offset": 0,
+      "estimatedTotalHits": 0,
+      "semanticHitCount": 0
+    }
+    "###);
+}

meilisearch/tests/vector/settings.rs

@@ -0,0 +1,228 @@
+use meili_snap::{json_string, snapshot};
+
+use crate::common::{GetAllDocumentsOptions, Server};
+use crate::json;
+use crate::vector::generate_default_user_provided_documents;
+
+#[actix_rt::test]
+async fn update_embedder() {
+    let server = Server::new().await;
+    let index = server.index("doggo");
+    let (value, code) = server.set_features(json!({"vectorStore": true})).await;
+    snapshot!(code, @"200 OK");
+    snapshot!(value, @r###"
+    {
+      "vectorStore": true,
+      "metrics": false,
+      "logsRoute": false
+    }
+    "###);
+
+    let (response, code) = index
+        .update_settings(json!({
+          "embedders": { "manual": {}},
+        }))
+        .await;
+    snapshot!(code, @"202 Accepted");
+    server.wait_task(response.uid()).await;
+
+    let (response, code) = index
+        .update_settings(json!({
+          "embedders": {
+              "manual": {
+                  "source": "userProvided",
+                  "dimensions": 2,
+              }
+          },
+        }))
+        .await;
+    snapshot!(code, @"202 Accepted");
+
+    let ret = server.wait_task(response.uid()).await;
+    snapshot!(ret, @r###"
+    {
+      "uid": 1,
+      "indexUid": "doggo",
+      "status": "succeeded",
+      "type": "settingsUpdate",
+      "canceledBy": null,
+      "details": {
+        "embedders": {
+          "manual": {
+            "source": "userProvided",
+            "dimensions": 2
+          }
+        }
+      },
+      "error": null,
+      "duration": "[duration]",
+      "enqueuedAt": "[date]",
+      "startedAt": "[date]",
+      "finishedAt": "[date]"
+    }
+    "###);
+}
+
+#[actix_rt::test]
+async fn reset_embedder_documents() {
+    let server = Server::new().await;
+    let index = generate_default_user_provided_documents(&server).await;
+
+    let (response, code) = index.delete_settings().await;
+    snapshot!(code, @"202 Accepted");
+    server.wait_task(response.uid()).await;
+
+    // Make sure the documents are still present
+    let (documents, _code) = index
+        .get_all_documents(GetAllDocumentsOptions {
+            limit: None,
+            offset: None,
+            retrieve_vectors: false,
+            fields: None,
+        })
+        .await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "results": [
+        {
+          "id": 0,
+          "name": "kefir"
+        },
+        {
+          "id": 1,
+          "name": "echo"
+        },
+        {
+          "id": 2,
+          "name": "billou"
+        },
+        {
+          "id": 3,
+          "name": "intel"
+        },
+        {
+          "id": 4,
+          "name": "max"
+        }
+      ],
+      "offset": 0,
+      "limit": 20,
+      "total": 5
+    }
+    "###);
+
+    // Make sure we are still able to retrieve their vectors
+    let (documents, _code) = index
+        .get_all_documents(GetAllDocumentsOptions { retrieve_vectors: true, ..Default::default() })
+        .await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "results": [
+        {
+          "id": 0,
+          "name": "kefir",
+          "_vectors": {
+            "manual": {
+              "embeddings": [
+                [
+                  0.0,
+                  0.0,
+                  0.0
+                ]
+              ],
+              "regenerate": false
+            }
+          }
+        },
+        {
+          "id": 1,
+          "name": "echo",
+          "_vectors": {
+            "manual": {
+              "embeddings": [
+                [
+                  1.0,
+                  1.0,
+                  1.0
+                ]
+              ],
+              "regenerate": false
+            }
+          }
+        },
+        {
+          "id": 2,
+          "name": "billou",
+          "_vectors": {
+            "manual": {
+              "embeddings": [
+                [
+                  2.0,
+                  2.0,
+                  2.0
+                ],
+                [
+                  2.0,
+                  2.0,
+                  3.0
+                ]
+              ],
+              "regenerate": false
+            }
+          }
+        },
+        {
+          "id": 3,
+          "name": "intel",
+          "_vectors": {
+            "manual": {
+              "embeddings": [
+                [
+                  3.0,
+                  3.0,
+                  3.0
+                ]
+              ],
+              "regenerate": false
+            }
+          }
+        },
+        {
+          "id": 4,
+          "name": "max",
+          "_vectors": {
+            "manual": {
+              "embeddings": [
+                [
+                  4.0,
+                  4.0,
+                  4.0
+                ],
+                [
+                  4.0,
+                  4.0,
+                  5.0
+                ]
+              ],
+              "regenerate": false
+            }
+          }
+        }
+      ],
+      "offset": 0,
+      "limit": 20,
+      "total": 5
+    }
+    "###);
+
+    // Make sure the arroy DB has been cleared
+    let (documents, _code) = index.search_post(json!({ "vector": [1, 1, 1] })).await;
+    snapshot!(json_string!(documents), @r###"
+    {
+      "message": "Cannot find embedder with name `default`.",
+      "code": "invalid_embedder",
+      "type": "invalid_request",
+      "link": "https://docs.meilisearch.com/errors#invalid_embedder"
+    }
+    "###);
+}

milli/Cargo.toml

@@ -44,7 +44,7 @@ once_cell = "1.19.0"
 ordered-float = "4.2.0"
 rand_pcg = { version = "0.3.1", features = ["serde1"] }
 rayon = "1.8.0"
-roaring = "0.10.2"
+roaring = { version = "0.10.2", features = ["serde"] }
 rstar = { version = "0.11.0", features = ["serde"] }
 serde = { version = "1.0.195", features = ["derive"] }
 serde_json = { version = "1.0.111", features = ["preserve_order"] }
@@ -71,10 +71,10 @@ csv = "1.3.0"
 candle-core = { version = "0.4.1" }
 candle-transformers = { version = "0.4.1" }
 candle-nn = { version = "0.4.1" }
-tokenizers = { git = "https://github.com/huggingface/tokenizers.git", tag = "v0.15.2", version = "0.15.2", default_features = false, features = [
+tokenizers = { git = "https://github.com/huggingface/tokenizers.git", tag = "v0.15.2", version = "0.15.2", default-features = false, features = [
     "onig",
 ] }
-hf-hub = { git = "https://github.com/dureuill/hf-hub.git", branch = "rust_tls", default_features = false, features = [
+hf-hub = { git = "https://github.com/dureuill/hf-hub.git", branch = "rust_tls", default-features = false, features = [
     "online",
 ] }
 tiktoken-rs = "0.5.8"

milli/src/fieldids_weights_map.rs

@@ -4,6 +4,7 @@ use std::collections::HashMap;
 
 use serde::{Deserialize, Serialize};
 
+use crate::vector::parsed_vectors::RESERVED_VECTORS_FIELD_NAME;
 use crate::{FieldId, FieldsIdsMap, Weight};
 
 #[derive(Debug, Default, Serialize, Deserialize)]
@@ -23,7 +24,13 @@ impl FieldidsWeightsMap {
     /// Should only be called in the case there are NO searchable attributes.
     /// All the fields will be inserted in the order of the fields ids map with a weight of 0.
     pub fn from_field_id_map_without_searchable(fid_map: &FieldsIdsMap) -> Self {
-        FieldidsWeightsMap { map: fid_map.ids().map(|fid| (fid, 0)).collect() }
+        FieldidsWeightsMap {
+            map: fid_map
+                .iter()
+                .filter(|(_fid, name)| !crate::is_faceted_by(name, RESERVED_VECTORS_FIELD_NAME))
+                .map(|(fid, _name)| (fid, 0))
+                .collect(),
+        }
     }
 
     /// Removes a field id from the map, returning the associated weight previously in the map.

milli/src/fields_ids_map.rs

@@ -41,6 +41,16 @@ impl FieldsIdsMap {
         }
     }
 
+    /// Get the ids of a field and all its nested fields based on its name.
+    pub fn nested_ids(&self, name: &str) -> Vec<FieldId> {
+        self.names_ids
+            .range(name.to_string()..)
+            .take_while(|(key, _)| key.starts_with(name))
+            .filter(|(key, _)| crate::is_faceted_by(key, name))
+            .map(|(_name, id)| *id)
+            .collect()
+    }
+
     /// Get the id of a field based on its name.
     pub fn id(&self, name: &str) -> Option<FieldId> {
         self.names_ids.get(name).copied()
@@ -126,4 +136,32 @@ mod tests {
         assert_eq!(iter.next(), Some((3, "title")));
         assert_eq!(iter.next(), None);
     }
+
+    #[test]
+    fn nested_fields() {
+        let mut map = FieldsIdsMap::new();
+
+        assert_eq!(map.insert("id"), Some(0));
+        assert_eq!(map.insert("doggo"), Some(1));
+        assert_eq!(map.insert("doggo.name"), Some(2));
+        assert_eq!(map.insert("doggolution"), Some(3));
+        assert_eq!(map.insert("doggo.breed.name"), Some(4));
+        assert_eq!(map.insert("description"), Some(5));
+
+        insta::assert_debug_snapshot!(map.nested_ids("doggo"), @r###"
+        [
+            1,
+            4,
+            2,
+        ]
+        "###);
+
+        insta::assert_debug_snapshot!(map.nested_ids("doggo.breed"), @r###"
+        [
+            4,
+        ]
+        "###);
+
+        insta::assert_debug_snapshot!(map.nested_ids("_vector"), @"[]");
+    }
 }

milli/src/index.rs

@@ -9,6 +9,7 @@ use heed::types::*;
 use heed::{CompactionOption, Database, RoTxn, RwTxn, Unspecified};
 use roaring::RoaringBitmap;
 use rstar::RTree;
+use serde::{Deserialize, Serialize};
 use time::OffsetDateTime;
 
 use crate::documents::PrimaryKey;
@@ -23,6 +24,7 @@ use crate::heed_codec::{
 };
 use crate::order_by_map::OrderByMap;
 use crate::proximity::ProximityPrecision;
+use crate::vector::parsed_vectors::RESERVED_VECTORS_FIELD_NAME;
 use crate::vector::{Embedding, EmbeddingConfig};
 use crate::{
     default_criteria, CboRoaringBitmapCodec, Criterion, DocumentId, ExternalDocumentsIds,
@@ -644,6 +646,7 @@ impl Index {
         &self,
         wtxn: &mut RwTxn,
         user_fields: &[&str],
+        non_searchable_fields_ids: &[FieldId],
         fields_ids_map: &FieldsIdsMap,
     ) -> Result<()> {
         // We can write the user defined searchable fields as-is.
@@ -662,6 +665,7 @@ impl Index {
             for (weight, user_field) in user_fields.iter().enumerate() {
                 if crate::is_faceted_by(field_from_map, user_field)
                     && !real_fields.contains(&field_from_map)
+                    && !non_searchable_fields_ids.contains(&id)
                 {
                     real_fields.push(field_from_map);
 
@@ -708,6 +712,7 @@ impl Index {
                 Ok(self
                     .fields_ids_map(rtxn)?
                     .names()
+                    .filter(|name| !crate::is_faceted_by(name, RESERVED_VECTORS_FIELD_NAME))
                     .map(|field| Cow::Owned(field.to_string()))
                     .collect())
             })
@@ -1568,12 +1573,16 @@ impl Index {
         Ok(script_language)
     }
 
+    /// Put the embedding configs:
+    /// 1. The name of the embedder
+    /// 2. The configuration option for this embedder
+    /// 3. The list of documents with a user provided embedding
     pub(crate) fn put_embedding_configs(
         &self,
         wtxn: &mut RwTxn<'_>,
-        configs: Vec<(String, EmbeddingConfig)>,
+        configs: Vec<IndexEmbeddingConfig>,
     ) -> heed::Result<()> {
-        self.main.remap_types::<Str, SerdeJson<Vec<(String, EmbeddingConfig)>>>().put(
+        self.main.remap_types::<Str, SerdeJson<Vec<IndexEmbeddingConfig>>>().put(
             wtxn,
             main_key::EMBEDDING_CONFIGS,
             &configs,
@@ -1584,13 +1593,10 @@ impl Index {
         self.main.remap_key_type::<Str>().delete(wtxn, main_key::EMBEDDING_CONFIGS)
     }
 
-    pub fn embedding_configs(
-        &self,
-        rtxn: &RoTxn<'_>,
-    ) -> Result<Vec<(String, crate::vector::EmbeddingConfig)>> {
+    pub fn embedding_configs(&self, rtxn: &RoTxn<'_>) -> Result<Vec<IndexEmbeddingConfig>> {
         Ok(self
             .main
-            .remap_types::<Str, SerdeJson<Vec<(String, EmbeddingConfig)>>>()
+            .remap_types::<Str, SerdeJson<Vec<IndexEmbeddingConfig>>>()
             .get(rtxn, main_key::EMBEDDING_CONFIGS)?
             .unwrap_or_default())
     }
@@ -1662,6 +1668,13 @@ impl Index {
     }
 }
 
+#[derive(Debug, Deserialize, Serialize)]
+pub struct IndexEmbeddingConfig {
+    pub name: String,
+    pub config: EmbeddingConfig,
+    pub user_provided: RoaringBitmap,
+}
+
 #[cfg(test)]
 pub(crate) mod tests {
     use std::collections::HashSet;
@@ -1669,15 +1682,17 @@ pub(crate) mod tests {
 
     use big_s::S;
     use heed::{EnvOpenOptions, RwTxn};
-    use maplit::hashset;
+    use maplit::{btreemap, hashset};
     use tempfile::TempDir;
 
     use crate::documents::DocumentsBatchReader;
     use crate::error::{Error, InternalError};
     use crate::index::{DEFAULT_MIN_WORD_LEN_ONE_TYPO, DEFAULT_MIN_WORD_LEN_TWO_TYPOS};
     use crate::update::{
-        self, IndexDocuments, IndexDocumentsConfig, IndexDocumentsMethod, IndexerConfig, Settings,
+        self, IndexDocuments, IndexDocumentsConfig, IndexDocumentsMethod, IndexerConfig, Setting,
+        Settings,
     };
+    use crate::vector::settings::{EmbedderSource, EmbeddingSettings};
     use crate::{db_snap, obkv_to_json, Filter, Index, Search, SearchResult};
 
     pub(crate) struct TempIndex {
@@ -2783,4 +2798,95 @@ pub(crate) mod tests {
         ]
         "###);
     }
+
+    #[test]
+    fn vectors_are_never_indexed_as_searchable_or_filterable() {
+        let index = TempIndex::new();
+
+        index
+            .add_documents(documents!([
+                { "id": 0, "_vectors": { "doggo": [2345] } },
+                { "id": 1, "_vectors": { "doggo": [6789] } },
+            ]))
+            .unwrap();
+
+        db_snap!(index, fields_ids_map, @r###"
+        0   id               |
+        1   _vectors         |
+        2   _vectors.doggo   |
+        "###);
+        db_snap!(index, searchable_fields, @r###"["id"]"###);
+        db_snap!(index, fieldids_weights_map, @r###"
+        fid weight
+        0   0   |
+        "###);
+
+        let rtxn = index.read_txn().unwrap();
+        let mut search = index.search(&rtxn);
+        let results = search.query("2345").execute().unwrap();
+        assert!(results.candidates.is_empty());
+        drop(rtxn);
+
+        index
+            .update_settings(|settings| {
+                settings.set_searchable_fields(vec![S("_vectors"), S("_vectors.doggo")]);
+                settings.set_filterable_fields(hashset![S("_vectors"), S("_vectors.doggo")]);
+            })
+            .unwrap();
+
+        db_snap!(index, fields_ids_map, @r###"
+        0   id               |
+        1   _vectors         |
+        2   _vectors.doggo   |
+        "###);
+        db_snap!(index, searchable_fields, @"[]");
+        db_snap!(index, fieldids_weights_map, @r###"
+        fid weight
+        "###);
+
+        let rtxn = index.read_txn().unwrap();
+        let mut search = index.search(&rtxn);
+        let results = search.query("2345").execute().unwrap();
+        assert!(results.candidates.is_empty());
+
+        let mut search = index.search(&rtxn);
+        let results = search
+            .filter(Filter::from_str("_vectors.doggo = 6789").unwrap().unwrap())
+            .execute()
+            .unwrap();
+        assert!(results.candidates.is_empty());
+
+        index
+            .update_settings(|settings| {
+                settings.set_embedder_settings(btreemap! {
+                    S("doggo") => Setting::Set(EmbeddingSettings {
+                        dimensions: Setting::Set(1),
+                        source: Setting::Set(EmbedderSource::UserProvided),
+                        ..EmbeddingSettings::default()}),
+                });
+            })
+            .unwrap();
+
+        db_snap!(index, fields_ids_map, @r###"
+        0   id               |
+        1   _vectors         |
+        2   _vectors.doggo   |
+        "###);
+        db_snap!(index, searchable_fields, @"[]");
+        db_snap!(index, fieldids_weights_map, @r###"
+        fid weight
+        "###);
+
+        let rtxn = index.read_txn().unwrap();
+        let mut search = index.search(&rtxn);
+        let results = search.query("2345").execute().unwrap();
+        assert!(results.candidates.is_empty());
+
+        let mut search = index.search(&rtxn);
+        let results = search
+            .filter(Filter::from_str("_vectors.doggo = 6789").unwrap().unwrap())
+            .execute()
+            .unwrap();
+        assert!(results.candidates.is_empty());
+    }
 }

milli/src/search/new/logger/visual.rs

@@ -22,7 +22,7 @@ pub enum SearchEvents {
     RankingRuleStartIteration { ranking_rule_idx: usize, universe_len: u64 },
     RankingRuleNextBucket { ranking_rule_idx: usize, universe_len: u64, bucket_len: u64 },
     RankingRuleSkipBucket { ranking_rule_idx: usize, bucket_len: u64 },
-    RankingRuleEndIteration { ranking_rule_idx: usize, universe_len: u64 },
+    RankingRuleEndIteration { ranking_rule_idx: usize },
     ExtendResults { new: Vec<u32> },
     ProximityGraph { graph: RankingRuleGraph<ProximityGraph> },
     ProximityPaths { paths: Vec<Vec<Interned<ProximityCondition>>> },
@@ -123,12 +123,9 @@ impl SearchLogger<QueryGraph> for VisualSearchLogger {
         &mut self,
         ranking_rule_idx: usize,
         _ranking_rule: &dyn RankingRule<QueryGraph>,
-        universe: &RoaringBitmap,
+        _universe: &RoaringBitmap,
     ) {
-        self.events.push(SearchEvents::RankingRuleEndIteration {
-            ranking_rule_idx,
-            universe_len: universe.len(),
-        });
+        self.events.push(SearchEvents::RankingRuleEndIteration { ranking_rule_idx });
         self.location.pop();
     }
     fn add_to_results(&mut self, docids: &[u32]) {
@@ -326,7 +323,7 @@ impl<'ctx> DetailedLoggerFinish<'ctx> {
                 assert!(ranking_rule_idx == self.rr_action_counter.len() - 1);
                 self.write_skip_bucket(bucket_len)?;
             }
-            SearchEvents::RankingRuleEndIteration { ranking_rule_idx, universe_len: _ } => {
+            SearchEvents::RankingRuleEndIteration { ranking_rule_idx } => {
                 assert!(ranking_rule_idx == self.rr_action_counter.len() - 1);
                 self.write_end_iteration()?;
             }

milli/src/search/new/tests/snapshots/milli__search__new__tests__attribute_fid__attribute_fid_ngrams.snap

@@ -1,244 +0,0 @@
----
-source: milli/src/search/new/tests/attribute_fid.rs
-expression: "format!(\"{document_ids_scores:#?}\")"
----
-[
-    (
-        2,
-        [
-            Fid(
-                Rank {
-                    rank: 19,
-                    max_rank: 19,
-                },
-            ),
-            Position(
-                Rank {
-                    rank: 91,
-                    max_rank: 91,
-                },
-            ),
-        ],
-    ),
-    (
-        6,
-        [
-            Fid(
-                Rank {
-                    rank: 15,
-                    max_rank: 19,
-                },
-            ),
-            Position(
-                Rank {
-                    rank: 81,
-                    max_rank: 91,
-                },
-            ),
-        ],
-    ),
-    (
-        5,
-        [
-            Fid(
-                Rank {
-                    rank: 14,
-                    max_rank: 19,
-                },
-            ),
-            Position(
-                Rank {
-                    rank: 79,
-                    max_rank: 91,
-                },
-            ),
-        ],
-    ),
-    (
-        4,
-        [
-            Fid(
-                Rank {
-                    rank: 13,
-                    max_rank: 19,
-                },
-            ),
-            Position(
-                Rank {
-                    rank: 77,
-                    max_rank: 91,
-                },
-            ),
-        ],
-    ),
-    (
-        3,
-        [
-            Fid(
-                Rank {
-                    rank: 12,
-                    max_rank: 19,
-                },
-            ),
-            Position(
-                Rank {
-                    rank: 83,
-                    max_rank: 91,
-                },
-            ),
-        ],
-    ),
-    (
-        9,
-        [
-            Fid(
-                Rank {
-                    rank: 11,
-                    max_rank: 19,
-                },
-            ),
-            Position(
-                Rank {
-                    rank: 75,
-                    max_rank: 91,
-                },
-            ),
-        ],
-    ),
-    (
-        8,
-        [
-            Fid(
-                Rank {
-                    rank: 10,
-                    max_rank: 19,
-                },
-            ),
-            Position(
-                Rank {
-                    rank: 79,
-                    max_rank: 91,
-                },
-            ),
-        ],
-    ),
-    (
-        7,
-        [
-            Fid(
-                Rank {
-                    rank: 10,
-                    max_rank: 19,
-                },
-            ),
-            Position(
-                Rank {
-                    rank: 73,
-                    max_rank: 91,
-                },
-            ),
-        ],
-    ),
-    (
-        11,
-        [
-            Fid(
-                Rank {
-                    rank: 7,
-                    max_rank: 19,
-                },
-            ),
-            Position(
-                Rank {
-                    rank: 77,
-                    max_rank: 91,
-                },
-            ),
-        ],
-    ),
-    (
-        10,
-        [
-            Fid(
-                Rank {
-                    rank: 6,
-                    max_rank: 19,
-                },
-            ),
-            Position(
-                Rank {
-                    rank: 81,
-                    max_rank: 91,
-                },
-            ),
-        ],
-    ),
-    (
-        13,
-        [
-            Fid(
-                Rank {
-                    rank: 6,
-                    max_rank: 19,
-                },
-            ),
-            Position(
-                Rank {
-                    rank: 81,
-                    max_rank: 91,
-                },
-            ),
-        ],
-    ),
-    (
-        12,
-        [
-            Fid(
-                Rank {
-                    rank: 6,
-                    max_rank: 19,
-                },
-            ),
-            Position(
-                Rank {
-                    rank: 78,
-                    max_rank: 91,
-                },
-            ),
-        ],
-    ),
-    (
-        14,
-        [
-            Fid(
-                Rank {
-                    rank: 5,
-                    max_rank: 19,
-                },
-            ),
-            Position(
-                Rank {
-                    rank: 75,
-                    max_rank: 91,
-                },
-            ),
-        ],
-    ),
-    (
-        0,
-        [
-            Fid(
-                Rank {
-                    rank: 1,
-                    max_rank: 19,
-                },
-            ),
-            Position(
-                Rank {
-                    rank: 91,
-                    max_rank: 91,
-                },
-            ),
-        ],
-    ),
-]

milli/src/snapshots/index.rs/initial_field_distribution/1/field_distribution.snap

@@ -1,7 +0,0 @@
----
-source: milli/src/index.rs
----
-age              1      |
-id               2      |
-name             2      |
-

milli/src/snapshots/index.rs/initial_field_distribution/field_distribution.snap

@@ -1,7 +0,0 @@
----
-source: milli/src/index.rs
----
-age              1      |
-id               2      |
-name             2      |
-

milli/src/update/clear_documents.rs

@@ -64,6 +64,13 @@ impl<'t, 'i> ClearDocuments<'t, 'i> {
         self.index.delete_geo_rtree(self.wtxn)?;
         self.index.delete_geo_faceted_documents_ids(self.wtxn)?;
 
+        // Remove all user-provided bits from the configs
+        let mut configs = self.index.embedding_configs(self.wtxn)?;
+        for config in configs.iter_mut() {
+            config.user_provided.clear();
+        }
+        self.index.put_embedding_configs(self.wtxn, configs)?;
+
         // Clear the other databases.
         external_documents_ids.clear(self.wtxn)?;
         word_docids.clear(self.wtxn)?;

milli/src/update/index_documents/extract/extract_vector_points.rs

@@ -8,18 +8,19 @@ use std::sync::Arc;
 
 use bytemuck::cast_slice;
 use grenad::Writer;
-use itertools::EitherOrBoth;
 use ordered_float::OrderedFloat;
+use roaring::RoaringBitmap;
 use serde_json::Value;
 
 use super::helpers::{create_writer, writer_into_reader, GrenadParameters};
+use crate::index::IndexEmbeddingConfig;
 use crate::prompt::Prompt;
 use crate::update::del_add::{DelAdd, KvReaderDelAdd, KvWriterDelAdd};
-use crate::update::index_documents::helpers::try_split_at;
 use crate::update::settings::InnerIndexSettingsDiff;
-use crate::vector::parsed_vectors::{ParsedVectorsDiff, RESERVED_VECTORS_FIELD_NAME};
+use crate::vector::parsed_vectors::{ParsedVectorsDiff, VectorState, RESERVED_VECTORS_FIELD_NAME};
+use crate::vector::settings::{EmbedderAction, ReindexAction};
 use crate::vector::Embedder;
-use crate::{DocumentId, Result, ThreadPoolNoAbort};
+use crate::{try_split_array_at, DocumentId, FieldId, FieldsIdsMap, Result, ThreadPoolNoAbort};
 
 /// The length of the elements that are always in the buffer when inserting new values.
 const TRUNCATE_SIZE: usize = size_of::<DocumentId>();
@@ -35,6 +36,8 @@ pub struct ExtractedVectorPoints {
     // embedder
     pub embedder_name: String,
     pub embedder: Arc<Embedder>,
+    pub add_to_user_provided: RoaringBitmap,
+    pub remove_from_user_provided: RoaringBitmap,
 }
 
 enum VectorStateDelta {
@@ -42,12 +45,7 @@ enum VectorStateDelta {
     // Remove all vectors, generated or manual, from this document
     NowRemoved,
 
-    // Add the manually specified vectors, passed in the other grenad
-    // Remove any previously generated vectors
-    // Note: changing the value of the manually specified vector **should not record** this delta
-    WasGeneratedNowManual(Vec<Vec<f32>>),
-
-    ManualDelta(Vec<Vec<f32>>, Vec<Vec<f32>>),
+    NowManual(Vec<Vec<f32>>),
 
     // Add the vector computed from the specified prompt
     // Remove any previous vector
@@ -56,14 +54,12 @@ enum VectorStateDelta {
 }
 
 impl VectorStateDelta {
-    fn into_values(self) -> (bool, String, (Vec<Vec<f32>>, Vec<Vec<f32>>)) {
+    fn into_values(self) -> (bool, String, Vec<Vec<f32>>) {
         match self {
             VectorStateDelta::NoChange => Default::default(),
             VectorStateDelta::NowRemoved => (true, Default::default(), Default::default()),
-            VectorStateDelta::WasGeneratedNowManual(add) => {
-                (true, Default::default(), (Default::default(), add))
-            }
-            VectorStateDelta::ManualDelta(del, add) => (false, Default::default(), (del, add)),
+            // We always delete the previous vectors
+            VectorStateDelta::NowManual(add) => (true, Default::default(), add),
             VectorStateDelta::NowGenerated(prompt) => (true, prompt, Default::default()),
         }
     }
@@ -74,12 +70,27 @@ struct EmbedderVectorExtractor {
     embedder: Arc<Embedder>,
     prompt: Arc<Prompt>,
 
-    // (docid, _index) -> KvWriterDelAdd -> Vector
-    manual_vectors_writer: Writer<BufWriter<File>>,
     // (docid) -> (prompt)
     prompts_writer: Writer<BufWriter<File>>,
     // (docid) -> ()
     remove_vectors_writer: Writer<BufWriter<File>>,
+    // (docid, _index) -> KvWriterDelAdd -> Vector
+    manual_vectors_writer: Writer<BufWriter<File>>,
+    // The docids of the documents that contains a user defined embedding
+    add_to_user_provided: RoaringBitmap,
+
+    action: ExtractionAction,
+}
+
+struct DocumentOperation {
+    // The docids of the documents that contains an auto-generated embedding
+    remove_from_user_provided: RoaringBitmap,
+}
+
+enum ExtractionAction {
+    SettingsFullReindex,
+    SettingsRegeneratePrompts { old_prompt: Arc<Prompt> },
+    DocumentOperation(DocumentOperation),
 }
 
 /// Extracts the embedding vector contained in each document under the `_vectors` field.
@@ -89,6 +100,7 @@ struct EmbedderVectorExtractor {
 pub fn extract_vector_points<R: io::Read + io::Seek>(
     obkv_documents: grenad::Reader<R>,
     indexer: GrenadParameters,
+    embedders_configs: &[IndexEmbeddingConfig],
     settings_diff: &InnerIndexSettingsDiff,
 ) -> Result<Vec<ExtractedVectorPoints>> {
     let reindex_vectors = settings_diff.reindex_vectors();
@@ -97,153 +109,207 @@ pub fn extract_vector_points<R: io::Read + io::Seek>(
     let new_fields_ids_map = &settings_diff.new.fields_ids_map;
     // the vector field id may have changed
     let old_vectors_fid = old_fields_ids_map.id(RESERVED_VECTORS_FIELD_NAME);
-    // filter the old vector fid if the settings has been changed forcing reindexing.
-    let old_vectors_fid = old_vectors_fid.filter(|_| !reindex_vectors);
 
     let new_vectors_fid = new_fields_ids_map.id(RESERVED_VECTORS_FIELD_NAME);
 
     let mut extractors = Vec::new();
-    for (embedder_name, (embedder, prompt)) in
-        settings_diff.new.embedding_configs.clone().into_iter()
-    {
-        // (docid, _index) -> KvWriterDelAdd -> Vector
-        let manual_vectors_writer = create_writer(
-            indexer.chunk_compression_type,
-            indexer.chunk_compression_level,
-            tempfile::tempfile()?,
-        );
 
-        // (docid) -> (prompt)
-        let prompts_writer = create_writer(
-            indexer.chunk_compression_type,
-            indexer.chunk_compression_level,
-            tempfile::tempfile()?,
-        );
+    let mut configs = settings_diff.new.embedding_configs.clone().into_inner();
+    let old_configs = &settings_diff.old.embedding_configs;
 
-        // (docid) -> ()
-        let remove_vectors_writer = create_writer(
-            indexer.chunk_compression_type,
-            indexer.chunk_compression_level,
-            tempfile::tempfile()?,
-        );
+    if reindex_vectors {
+        for (name, action) in settings_diff.embedding_config_updates.iter() {
+            match action {
+                EmbedderAction::WriteBackToDocuments(_) => continue, // already deleted
+                EmbedderAction::Reindex(action) => {
+                    let Some((embedder_name, (embedder, prompt))) = configs.remove_entry(name)
+                    else {
+                        tracing::error!(embedder = name, "Requested embedder config not found");
+                        continue;
+                    };
 
-        extractors.push(EmbedderVectorExtractor {
-            embedder_name,
-            embedder,
-            prompt,
-            manual_vectors_writer,
-            prompts_writer,
-            remove_vectors_writer,
-        });
+                    // (docid, _index) -> KvWriterDelAdd -> Vector
+                    let manual_vectors_writer = create_writer(
+                        indexer.chunk_compression_type,
+                        indexer.chunk_compression_level,
+                        tempfile::tempfile()?,
+                    );
+
+                    // (docid) -> (prompt)
+                    let prompts_writer = create_writer(
+                        indexer.chunk_compression_type,
+                        indexer.chunk_compression_level,
+                        tempfile::tempfile()?,
+                    );
+
+                    // (docid) -> ()
+                    let remove_vectors_writer = create_writer(
+                        indexer.chunk_compression_type,
+                        indexer.chunk_compression_level,
+                        tempfile::tempfile()?,
+                    );
+
+                    let action = match action {
+                        ReindexAction::FullReindex => ExtractionAction::SettingsFullReindex,
+                        ReindexAction::RegeneratePrompts => {
+                            let Some((_, old_prompt)) = old_configs.get(name) else {
+                                tracing::error!(embedder = name, "Old embedder config not found");
+                                continue;
+                            };
+
+                            ExtractionAction::SettingsRegeneratePrompts { old_prompt }
+                        }
+                    };
+
+                    extractors.push(EmbedderVectorExtractor {
+                        embedder_name,
+                        embedder,
+                        prompt,
+                        prompts_writer,
+                        remove_vectors_writer,
+                        manual_vectors_writer,
+                        add_to_user_provided: RoaringBitmap::new(),
+                        action,
+                    });
+                }
+            }
+        }
+    } else {
+        // document operation
+
+        for (embedder_name, (embedder, prompt)) in configs.into_iter() {
+            // (docid, _index) -> KvWriterDelAdd -> Vector
+            let manual_vectors_writer = create_writer(
+                indexer.chunk_compression_type,
+                indexer.chunk_compression_level,
+                tempfile::tempfile()?,
+            );
+
+            // (docid) -> (prompt)
+            let prompts_writer = create_writer(
+                indexer.chunk_compression_type,
+                indexer.chunk_compression_level,
+                tempfile::tempfile()?,
+            );
+
+            // (docid) -> ()
+            let remove_vectors_writer = create_writer(
+                indexer.chunk_compression_type,
+                indexer.chunk_compression_level,
+                tempfile::tempfile()?,
+            );
+
+            extractors.push(EmbedderVectorExtractor {
+                embedder_name,
+                embedder,
+                prompt,
+                prompts_writer,
+                remove_vectors_writer,
+                manual_vectors_writer,
+                add_to_user_provided: RoaringBitmap::new(),
+                action: ExtractionAction::DocumentOperation(DocumentOperation {
+                    remove_from_user_provided: RoaringBitmap::new(),
+                }),
+            });
+        }
     }
 
     let mut key_buffer = Vec::new();
     let mut cursor = obkv_documents.into_cursor()?;
     while let Some((key, value)) = cursor.move_on_next()? {
         // this must always be serialized as (docid, external_docid);
+        const SIZE_OF_DOCUMENTID: usize = std::mem::size_of::<DocumentId>();
         let (docid_bytes, external_id_bytes) =
-            try_split_at(key, std::mem::size_of::<DocumentId>()).unwrap();
+            try_split_array_at::<u8, SIZE_OF_DOCUMENTID>(key).unwrap();
         debug_assert!(from_utf8(external_id_bytes).is_ok());
+        let docid = DocumentId::from_be_bytes(docid_bytes);
 
         let obkv = obkv::KvReader::new(value);
         key_buffer.clear();
-        key_buffer.extend_from_slice(docid_bytes);
+        key_buffer.extend_from_slice(docid_bytes.as_slice());
 
         // since we only need the primary key when we throw an error we create this getter to
         // lazily get it when needed
         let document_id = || -> Value { from_utf8(external_id_bytes).unwrap().into() };
 
-        let mut parsed_vectors = ParsedVectorsDiff::new(obkv, old_vectors_fid, new_vectors_fid)
-            .map_err(|error| error.to_crate_error(document_id().to_string()))?;
+        let mut parsed_vectors = ParsedVectorsDiff::new(
+            docid,
+            embedders_configs,
+            obkv,
+            old_vectors_fid,
+            new_vectors_fid,
+        )
+        .map_err(|error| error.to_crate_error(document_id().to_string()))?;
 
         for EmbedderVectorExtractor {
             embedder_name,
             embedder: _,
             prompt,
-            manual_vectors_writer,
             prompts_writer,
             remove_vectors_writer,
+            manual_vectors_writer,
+            add_to_user_provided,
+            action,
         } in extractors.iter_mut()
         {
-            let delta = match parsed_vectors.remove(embedder_name) {
-                (Some(old), Some(new)) => {
-                    // no autogeneration
-                    let del_vectors = old.into_array_of_vectors();
-                    let add_vectors = new.into_array_of_vectors();
-
-                    if add_vectors.len() > usize::from(u8::MAX) {
-                        return Err(crate::Error::UserError(crate::UserError::TooManyVectors(
-                            document_id().to_string(),
-                            add_vectors.len(),
-                        )));
-                    }
-
-                    VectorStateDelta::ManualDelta(del_vectors, add_vectors)
-                }
-                (Some(_old), None) => {
-                    // Do we keep this document?
-                    let document_is_kept = obkv
-                        .iter()
-                        .map(|(_, deladd)| KvReaderDelAdd::new(deladd))
-                        .any(|deladd| deladd.get(DelAdd::Addition).is_some());
-                    if document_is_kept {
-                        // becomes autogenerated
-                        VectorStateDelta::NowGenerated(prompt.render(
-                            obkv,
-                            DelAdd::Addition,
-                            new_fields_ids_map,
-                        )?)
-                    } else {
-                        VectorStateDelta::NowRemoved
-                    }
-                }
-                (None, Some(new)) => {
-                    // was possibly autogenerated, remove all vectors for that document
-                    let add_vectors = new.into_array_of_vectors();
-                    if add_vectors.len() > usize::from(u8::MAX) {
-                        return Err(crate::Error::UserError(crate::UserError::TooManyVectors(
-                            document_id().to_string(),
-                            add_vectors.len(),
-                        )));
-                    }
-
-                    VectorStateDelta::WasGeneratedNowManual(add_vectors)
-                }
-                (None, None) => {
-                    // Do we keep this document?
-                    let document_is_kept = obkv
-                        .iter()
-                        .map(|(_, deladd)| KvReaderDelAdd::new(deladd))
-                        .any(|deladd| deladd.get(DelAdd::Addition).is_some());
-
-                    if document_is_kept {
-                        // Don't give up if the old prompt was failing
-                        let old_prompt = Some(&prompt)
-                            // TODO: this filter works because we erase the vec database when a embedding setting changes.
-                            // When vector pipeline will be optimized, this should be removed.
-                            .filter(|_| !settings_diff.reindex_vectors())
-                            .map(|p| {
-                                p.render(obkv, DelAdd::Deletion, old_fields_ids_map)
-                                    .unwrap_or_default()
-                            });
-                        let new_prompt =
-                            prompt.render(obkv, DelAdd::Addition, new_fields_ids_map)?;
-                        if old_prompt.as_ref() != Some(&new_prompt) {
-                            let old_prompt = old_prompt.unwrap_or_default();
-                            tracing::trace!(
-                                "🚀 Changing prompt from\n{old_prompt}\n===to===\n{new_prompt}"
-                            );
-                            VectorStateDelta::NowGenerated(new_prompt)
-                        } else {
-                            tracing::trace!("⏭️ Prompt unmodified, skipping");
-                            VectorStateDelta::NoChange
+            let (old, new) = parsed_vectors.remove(embedder_name);
+            let delta = match action {
+                ExtractionAction::SettingsFullReindex => match old {
+                    // A full reindex can be triggered either by:
+                    // 1. a new embedder
+                    // 2. an existing embedder changed so that it must regenerate all generated embeddings.
+                    // For a new embedder, there can be `_vectors.embedder` embeddings to add to the DB
+                    VectorState::Inline(vectors) => {
+                        if !vectors.must_regenerate() {
+                            add_to_user_provided.insert(docid);
                         }
+
+                        match vectors.into_array_of_vectors() {
+                            Some(add_vectors) => {
+                                if add_vectors.len() > usize::from(u8::MAX) {
+                                    return Err(crate::Error::UserError(
+                                        crate::UserError::TooManyVectors(
+                                            document_id().to_string(),
+                                            add_vectors.len(),
+                                        ),
+                                    ));
+                                }
+                                VectorStateDelta::NowManual(add_vectors)
+                            }
+                            None => VectorStateDelta::NoChange,
+                        }
+                    }
+                    // this happens only when an existing embedder changed. We cannot regenerate userProvided vectors
+                    VectorState::Manual => VectorStateDelta::NoChange,
+                    // generated vectors must be regenerated
+                    VectorState::Generated => regenerate_prompt(obkv, prompt, new_fields_ids_map)?,
+                },
+                // prompt regeneration is only triggered for existing embedders
+                ExtractionAction::SettingsRegeneratePrompts { old_prompt } => {
+                    if old.must_regenerate() {
+                        regenerate_if_prompt_changed(
+                            obkv,
+                            (old_prompt, prompt),
+                            (&old_fields_ids_map, &new_fields_ids_map),
+                        )?
                     } else {
-                        VectorStateDelta::NowRemoved
+                        // we can simply ignore user provided vectors as they are not regenerated and are
+                        // already in the DB since this is an existing embedder
+                        VectorStateDelta::NoChange
                     }
                 }
+                ExtractionAction::DocumentOperation(DocumentOperation {
+                    remove_from_user_provided,
+                }) => extract_vector_document_diff(
+                    docid,
+                    obkv,
+                    prompt,
+                    (add_to_user_provided, remove_from_user_provided),
+                    (old, new),
+                    (&old_fields_ids_map, &new_fields_ids_map),
+                    document_id,
+                )?,
             };
-
             // and we finally push the unique vectors into the writer
             push_vectors_diff(
                 remove_vectors_writer,
@@ -251,7 +317,6 @@ pub fn extract_vector_points<R: io::Read + io::Seek>(
                 manual_vectors_writer,
                 &mut key_buffer,
                 delta,
-                reindex_vectors,
             )?;
         }
     }
@@ -262,43 +327,185 @@ pub fn extract_vector_points<R: io::Read + io::Seek>(
         embedder_name,
         embedder,
         prompt: _,
-        manual_vectors_writer,
         prompts_writer,
         remove_vectors_writer,
+        action,
+        manual_vectors_writer,
+        add_to_user_provided,
     } in extractors
     {
-        results.push(ExtractedVectorPoints {
-            // docid, _index -> KvWriterDelAdd -> Vector
-            manual_vectors: writer_into_reader(manual_vectors_writer)?,
-            // docid -> ()
-            remove_vectors: writer_into_reader(remove_vectors_writer)?,
-            // docid -> prompt
-            prompts: writer_into_reader(prompts_writer)?,
+        let remove_from_user_provided =
+            if let ExtractionAction::DocumentOperation(DocumentOperation {
+                remove_from_user_provided,
+            }) = action
+            {
+                remove_from_user_provided
+            } else {
+                Default::default()
+            };
 
+        results.push(ExtractedVectorPoints {
+            manual_vectors: writer_into_reader(manual_vectors_writer)?,
+            remove_vectors: writer_into_reader(remove_vectors_writer)?,
+            prompts: writer_into_reader(prompts_writer)?,
             embedder,
             embedder_name,
+            add_to_user_provided,
+            remove_from_user_provided,
         })
     }
 
     Ok(results)
 }
 
-/// Computes the diff between both Del and Add numbers and
-/// only inserts the parts that differ in the sorter.
+fn extract_vector_document_diff(
+    docid: DocumentId,
+    obkv: obkv::KvReader<'_, FieldId>,
+    prompt: &Prompt,
+    (add_to_user_provided, remove_from_user_provided): (&mut RoaringBitmap, &mut RoaringBitmap),
+    (old, new): (VectorState, VectorState),
+    (old_fields_ids_map, new_fields_ids_map): (&FieldsIdsMap, &FieldsIdsMap),
+    document_id: impl Fn() -> Value,
+) -> Result<VectorStateDelta> {
+    match (old.must_regenerate(), new.must_regenerate()) {
+        (true, true) | (false, false) => {}
+        (true, false) => {
+            add_to_user_provided.insert(docid);
+        }
+        (false, true) => {
+            remove_from_user_provided.insert(docid);
+        }
+    }
+
+    let delta = match (old, new) {
+        // regardless of the previous state, if a document now contains inline _vectors, they must
+        // be extracted manually
+        (_old, VectorState::Inline(new)) => match new.into_array_of_vectors() {
+            Some(add_vectors) => {
+                if add_vectors.len() > usize::from(u8::MAX) {
+                    return Err(crate::Error::UserError(crate::UserError::TooManyVectors(
+                        document_id().to_string(),
+                        add_vectors.len(),
+                    )));
+                }
+
+                VectorStateDelta::NowManual(add_vectors)
+            }
+            None => VectorStateDelta::NoChange,
+        },
+        // no `_vectors` anywhere, we check for document removal and otherwise we regenerate the prompt if the
+        // document changed
+        (VectorState::Generated, VectorState::Generated) => {
+            // Do we keep this document?
+            let document_is_kept = obkv
+                .iter()
+                .map(|(_, deladd)| KvReaderDelAdd::new(deladd))
+                .any(|deladd| deladd.get(DelAdd::Addition).is_some());
+
+            if document_is_kept {
+                // Don't give up if the old prompt was failing
+                let old_prompt = Some(&prompt).map(|p| {
+                    p.render(obkv, DelAdd::Deletion, old_fields_ids_map).unwrap_or_default()
+                });
+                let new_prompt = prompt.render(obkv, DelAdd::Addition, new_fields_ids_map)?;
+                if old_prompt.as_ref() != Some(&new_prompt) {
+                    let old_prompt = old_prompt.unwrap_or_default();
+                    tracing::trace!(
+                        "🚀 Changing prompt from\n{old_prompt}\n===to===\n{new_prompt}"
+                    );
+                    VectorStateDelta::NowGenerated(new_prompt)
+                } else {
+                    tracing::trace!("⏭️ Prompt unmodified, skipping");
+                    VectorStateDelta::NoChange
+                }
+            } else {
+                VectorStateDelta::NowRemoved
+            }
+        }
+        // inline to the left is not supposed to be possible because the embedder is not new, so `_vectors` was removed from
+        // the previous version of the document.
+        // Manual -> Generated is also not possible without an Inline to the right (which is handled above)
+        // Generated -> Generated is handled above, so not possible
+        // As a result, this code is unreachable
+        (_not_generated, VectorState::Generated) => {
+            // Do we keep this document?
+            let document_is_kept = obkv
+                .iter()
+                .map(|(_, deladd)| KvReaderDelAdd::new(deladd))
+                .any(|deladd| deladd.get(DelAdd::Addition).is_some());
+            if document_is_kept {
+                // becomes autogenerated
+                VectorStateDelta::NowGenerated(prompt.render(
+                    obkv,
+                    DelAdd::Addition,
+                    new_fields_ids_map,
+                )?)
+            } else {
+                // make sure the document is always removed from user provided on removal
+                remove_from_user_provided.insert(docid);
+                VectorStateDelta::NowRemoved
+            }
+        }
+        // inline to the left is not possible because the embedder is not new, and so `_vectors` was removed from the previous
+        // version of the document.
+        // however the Rust type system cannot know that.
+        (_manual, VectorState::Manual) => {
+            // Do we keep this document?
+            let document_is_kept = obkv
+                .iter()
+                .map(|(_, deladd)| KvReaderDelAdd::new(deladd))
+                .any(|deladd| deladd.get(DelAdd::Addition).is_some());
+            if document_is_kept {
+                // if the new version of documents has the vectors in the DB,
+                // then they are user-provided and nothing possibly changed
+                VectorStateDelta::NoChange
+            } else {
+                // make sure the document is always removed from user provided on removal
+                remove_from_user_provided.insert(docid);
+                VectorStateDelta::NowRemoved
+            }
+        }
+    };
+
+    Ok(delta)
+}
+
+fn regenerate_if_prompt_changed(
+    obkv: obkv::KvReader<'_, FieldId>,
+    (old_prompt, new_prompt): (&Prompt, &Prompt),
+    (old_fields_ids_map, new_fields_ids_map): (&FieldsIdsMap, &FieldsIdsMap),
+) -> Result<VectorStateDelta> {
+    let old_prompt =
+        old_prompt.render(obkv, DelAdd::Deletion, old_fields_ids_map).unwrap_or(Default::default());
+    let new_prompt = new_prompt.render(obkv, DelAdd::Addition, new_fields_ids_map)?;
+
+    if new_prompt == old_prompt {
+        return Ok(VectorStateDelta::NoChange);
+    }
+    Ok(VectorStateDelta::NowGenerated(new_prompt))
+}
+
+fn regenerate_prompt(
+    obkv: obkv::KvReader<'_, FieldId>,
+    prompt: &Prompt,
+    new_fields_ids_map: &FieldsIdsMap,
+) -> Result<VectorStateDelta> {
+    let prompt = prompt.render(obkv, DelAdd::Addition, new_fields_ids_map)?;
+
+    Ok(VectorStateDelta::NowGenerated(prompt))
+}
+
+/// We cannot compute the diff between both Del and Add vectors.
+/// We'll push every vector and compute the difference later in TypedChunk.
 fn push_vectors_diff(
     remove_vectors_writer: &mut Writer<BufWriter<File>>,
     prompts_writer: &mut Writer<BufWriter<File>>,
     manual_vectors_writer: &mut Writer<BufWriter<File>>,
     key_buffer: &mut Vec<u8>,
     delta: VectorStateDelta,
-    reindex_vectors: bool,
 ) -> Result<()> {
-    let (must_remove, prompt, (mut del_vectors, mut add_vectors)) = delta.into_values();
-    if must_remove
-    // TODO: the below condition works because we erase the vec database when a embedding setting changes.
-    // When vector pipeline will be optimized, this should be removed.
-    && !reindex_vectors
-    {
+    let (must_remove, prompt, mut add_vectors) = delta.into_values();
+    if must_remove {
         key_buffer.truncate(TRUNCATE_SIZE);
         remove_vectors_writer.insert(&key_buffer, [])?;
     }
@@ -308,44 +515,22 @@ fn push_vectors_diff(
     }
 
     // We sort and dedup the vectors
-    del_vectors.sort_unstable_by(|a, b| compare_vectors(a, b));
     add_vectors.sort_unstable_by(|a, b| compare_vectors(a, b));
-    del_vectors.dedup_by(|a, b| compare_vectors(a, b).is_eq());
     add_vectors.dedup_by(|a, b| compare_vectors(a, b).is_eq());
 
-    let merged_vectors_iter =
-        itertools::merge_join_by(del_vectors, add_vectors, |del, add| compare_vectors(del, add));
-
     // insert vectors into the writer
-    for (i, eob) in merged_vectors_iter.into_iter().enumerate().take(u16::MAX as usize) {
+    for (i, vector) in add_vectors.into_iter().enumerate().take(u16::MAX as usize) {
         // Generate the key by extending the unique index to it.
         key_buffer.truncate(TRUNCATE_SIZE);
         let index = u16::try_from(i).unwrap();
         key_buffer.extend_from_slice(&index.to_be_bytes());
 
-        match eob {
-            EitherOrBoth::Both(_, _) => (), // no need to touch anything
-            EitherOrBoth::Left(vector) => {
-                // TODO: the below condition works because we erase the vec database when a embedding setting changes.
-                // When vector pipeline will be optimized, this should be removed.
-                if !reindex_vectors {
-                    // We insert only the Del part of the Obkv to inform
-                    // that we only want to remove all those vectors.
-                    let mut obkv = KvWriterDelAdd::memory();
-                    obkv.insert(DelAdd::Deletion, cast_slice(&vector))?;
-                    let bytes = obkv.into_inner()?;
-                    manual_vectors_writer.insert(&key_buffer, bytes)?;
-                }
-            }
-            EitherOrBoth::Right(vector) => {
-                // We insert only the Add part of the Obkv to inform
-                // that we only want to remove all those vectors.
-                let mut obkv = KvWriterDelAdd::memory();
-                obkv.insert(DelAdd::Addition, cast_slice(&vector))?;
-                let bytes = obkv.into_inner()?;
-                manual_vectors_writer.insert(&key_buffer, bytes)?;
-            }
-        }
+        // We insert only the Add part of the Obkv to inform
+        // that we only want to remove all those vectors.
+        let mut obkv = KvWriterDelAdd::memory();
+        obkv.insert(DelAdd::Addition, cast_slice(&vector))?;
+        let bytes = obkv.into_inner()?;
+        manual_vectors_writer.insert(&key_buffer, bytes)?;
     }
 
     Ok(())

milli/src/update/index_documents/extract/mod.rs

@@ -30,6 +30,7 @@ use self::extract_word_pair_proximity_docids::extract_word_pair_proximity_docids
 use self::extract_word_position_docids::extract_word_position_docids;
 use super::helpers::{as_cloneable_grenad, CursorClonableMmap, GrenadParameters};
 use super::{helpers, TypedChunk};
+use crate::index::IndexEmbeddingConfig;
 use crate::update::settings::InnerIndexSettingsDiff;
 use crate::{FieldId, Result, ThreadPoolNoAbortBuilder};
 
@@ -43,6 +44,7 @@ pub(crate) fn data_from_obkv_documents(
     indexer: GrenadParameters,
     lmdb_writer_sx: Sender<Result<TypedChunk>>,
     primary_key_id: FieldId,
+    embedders_configs: Arc<Vec<IndexEmbeddingConfig>>,
     settings_diff: Arc<InnerIndexSettingsDiff>,
     max_positions_per_attributes: Option<u32>,
 ) -> Result<()> {
@@ -55,6 +57,7 @@ pub(crate) fn data_from_obkv_documents(
                         original_documents_chunk,
                         indexer,
                         lmdb_writer_sx.clone(),
+                        embedders_configs.clone(),
                         settings_diff.clone(),
                     )
                 })
@@ -210,6 +213,7 @@ fn send_original_documents_data(
     original_documents_chunk: Result<grenad::Reader<BufReader<File>>>,
     indexer: GrenadParameters,
     lmdb_writer_sx: Sender<Result<TypedChunk>>,
+    embedders_configs: Arc<Vec<IndexEmbeddingConfig>>,
     settings_diff: Arc<InnerIndexSettingsDiff>,
 ) -> Result<()> {
     let original_documents_chunk =
@@ -226,11 +230,17 @@ fn send_original_documents_data(
 
     if index_vectors {
         let settings_diff = settings_diff.clone();
+        let embedders_configs = embedders_configs.clone();
 
         let original_documents_chunk = original_documents_chunk.clone();
         let lmdb_writer_sx = lmdb_writer_sx.clone();
         rayon::spawn(move || {
-            match extract_vector_points(original_documents_chunk.clone(), indexer, &settings_diff) {
+            match extract_vector_points(
+                original_documents_chunk.clone(),
+                indexer,
+                &embedders_configs,
+                &settings_diff,
+            ) {
                 Ok(extracted_vectors) => {
                     for ExtractedVectorPoints {
                         manual_vectors,
@@ -238,6 +248,8 @@ fn send_original_documents_data(
                         prompts,
                         embedder_name,
                         embedder,
+                        add_to_user_provided,
+                        remove_from_user_provided,
                     } in extracted_vectors
                     {
                         let embeddings = match extract_embeddings(
@@ -262,6 +274,8 @@ fn send_original_documents_data(
                                 expected_dimension: embedder.dimensions(),
                                 manual_vectors,
                                 embedder_name,
+                                add_to_user_provided,
+                                remove_from_user_provided,
                             }));
                         }
                     }

milli/src/update/index_documents/mod.rs

@@ -286,6 +286,7 @@ where
         settings_diff.new.recompute_searchables(self.wtxn, self.index)?;
 
         let settings_diff = Arc::new(settings_diff);
+        let embedders_configs = Arc::new(self.index.embedding_configs(self.wtxn)?);
 
         let backup_pool;
         let pool = match self.indexer_config.thread_pool {
@@ -399,6 +400,7 @@ where
                         pool_params,
                         lmdb_writer_sx.clone(),
                         primary_key_id,
+                        embedders_configs.clone(),
                         settings_diff_cloned,
                         max_positions_per_attributes,
                     )
@@ -501,6 +503,8 @@ where
                                 embeddings,
                                 manual_vectors,
                                 embedder_name,
+                                add_to_user_provided,
+                                remove_from_user_provided,
                             } => {
                                 dimension.insert(embedder_name.clone(), expected_dimension);
                                 TypedChunk::VectorPoints {
@@ -509,6 +513,8 @@ where
                                     expected_dimension,
                                     manual_vectors,
                                     embedder_name,
+                                    add_to_user_provided,
+                                    remove_from_user_provided,
                                 }
                             }
                             otherwise => otherwise,
@@ -781,6 +787,7 @@ mod tests {
     use super::*;
     use crate::documents::documents_batch_reader_from_objects;
     use crate::index::tests::TempIndex;
+    use crate::index::IndexEmbeddingConfig;
     use crate::search::TermsMatchingStrategy;
     use crate::update::Setting;
     use crate::{db_snap, Filter, Search};
@@ -2616,10 +2623,12 @@ mod tests {
 
         let rtxn = index.read_txn().unwrap();
         let mut embedding_configs = index.embedding_configs(&rtxn).unwrap();
-        let (embedder_name, embedder) = embedding_configs.pop().unwrap();
+        let IndexEmbeddingConfig { name: embedder_name, config: embedder, user_provided } =
+            embedding_configs.pop().unwrap();
+        insta::assert_snapshot!(embedder_name, @"manual");
+        insta::assert_debug_snapshot!(user_provided, @"RoaringBitmap<[0, 1, 2]>");
         let embedder =
             std::sync::Arc::new(crate::vector::Embedder::new(embedder.embedder_options).unwrap());
-        assert_eq!("manual", embedder_name);
         let res = index
             .search(&rtxn)
             .semantic(embedder_name, embedder, Some([0.0, 1.0, 2.0].to_vec()))

milli/src/update/index_documents/transform.rs

@@ -1,7 +1,7 @@
 use std::borrow::Cow;
 use std::collections::btree_map::Entry as BEntry;
 use std::collections::hash_map::Entry as HEntry;
-use std::collections::{HashMap, HashSet};
+use std::collections::{BTreeMap, HashMap, HashSet};
 use std::fs::File;
 use std::io::{Read, Seek};
 
@@ -27,6 +27,8 @@ use crate::update::del_add::{
 use crate::update::index_documents::GrenadParameters;
 use crate::update::settings::{InnerIndexSettings, InnerIndexSettingsDiff};
 use crate::update::{AvailableDocumentsIds, UpdateIndexingStep};
+use crate::vector::parsed_vectors::{ExplicitVectors, VectorOrArrayOfVectors};
+use crate::vector::settings::{EmbedderAction, WriteBackToDocuments};
 use crate::{
     is_faceted_by, FieldDistribution, FieldId, FieldIdMapMissingEntry, FieldsIdsMap, Index, Result,
 };
@@ -51,7 +53,6 @@ pub struct Transform<'a, 'i> {
     fields_ids_map: FieldsIdsMap,
 
     indexer_settings: &'a IndexerConfig,
-    pub autogenerate_docids: bool,
     pub index_documents_method: IndexDocumentsMethod,
     available_documents_ids: AvailableDocumentsIds,
 
@@ -105,7 +106,7 @@ impl<'a, 'i> Transform<'a, 'i> {
         index: &'i Index,
         indexer_settings: &'a IndexerConfig,
         index_documents_method: IndexDocumentsMethod,
-        autogenerate_docids: bool,
+        _autogenerate_docids: bool,
     ) -> Result<Self> {
         // We must choose the appropriate merge function for when two or more documents
         // with the same user id must be merged or fully replaced in the same batch.
@@ -139,7 +140,6 @@ impl<'a, 'i> Transform<'a, 'i> {
             index,
             fields_ids_map: index.fields_ids_map(wtxn)?,
             indexer_settings,
-            autogenerate_docids,
             available_documents_ids: AvailableDocumentsIds::from_documents_ids(&documents_ids),
             original_sorter,
             flattened_sorter,
@@ -808,13 +808,13 @@ impl<'a, 'i> Transform<'a, 'i> {
         let mut new_inner_settings = old_inner_settings.clone();
         new_inner_settings.fields_ids_map = fields_ids_map;
 
-        let embedding_configs_updated = false;
+        let embedding_config_updates = Default::default();
         let settings_update_only = false;
         let settings_diff = InnerIndexSettingsDiff::new(
             old_inner_settings,
             new_inner_settings,
             primary_key_id,
-            embedding_configs_updated,
+            embedding_config_updates,
             settings_update_only,
         );
 
@@ -835,10 +835,13 @@ impl<'a, 'i> Transform<'a, 'i> {
     /// Rebind the field_ids of the provided document to their values
     /// based on the field_ids_maps difference between the old and the new settings,
     /// then fill the provided buffers with delta documents using KvWritterDelAdd.
+    #[allow(clippy::too_many_arguments)] // need the vectors + fid, feel free to create a struct xo xo
     fn rebind_existing_document(
         old_obkv: KvReader<FieldId>,
         settings_diff: &InnerIndexSettingsDiff,
         modified_faceted_fields: &HashSet<String>,
+        mut injected_vectors: serde_json::Map<String, serde_json::Value>,
+        old_vectors_fid: Option<FieldId>,
         original_obkv_buffer: Option<&mut Vec<u8>>,
         flattened_obkv_buffer: Option<&mut Vec<u8>>,
     ) -> Result<()> {
@@ -861,9 +864,49 @@ impl<'a, 'i> Transform<'a, 'i> {
 
         // The operations that we must perform on the different fields.
         let mut operations = HashMap::new();
+        let mut error_seen = false;
 
         let mut obkv_writer = KvWriter::<_, FieldId>::memory();
-        for (id, val) in old_obkv.iter() {
+        'write_fid: for (id, val) in old_obkv.iter() {
+            if !injected_vectors.is_empty() {
+                'inject_vectors: {
+                    let Some(vectors_fid) = old_vectors_fid else { break 'inject_vectors };
+
+                    if id < vectors_fid {
+                        break 'inject_vectors;
+                    }
+
+                    let mut existing_vectors = if id == vectors_fid {
+                        let existing_vectors: std::result::Result<
+                            serde_json::Map<String, serde_json::Value>,
+                            serde_json::Error,
+                        > = serde_json::from_slice(val);
+
+                        match existing_vectors {
+                            Ok(existing_vectors) => existing_vectors,
+                            Err(error) => {
+                                if !error_seen {
+                                    tracing::error!(%error, "Unexpected `_vectors` field that is not a map. Treating as an empty map");
+                                    error_seen = true;
+                                }
+                                Default::default()
+                            }
+                        }
+                    } else {
+                        Default::default()
+                    };
+
+                    existing_vectors.append(&mut injected_vectors);
+
+                    operations.insert(vectors_fid, DelAddOperation::DeletionAndAddition);
+                    obkv_writer
+                        .insert(vectors_fid, serde_json::to_vec(&existing_vectors).unwrap())?;
+                    if id == vectors_fid {
+                        continue 'write_fid;
+                    }
+                }
+            }
+
             if is_primary_key(id) || necessary_faceted_field(id) || reindex_vectors {
                 operations.insert(id, DelAddOperation::DeletionAndAddition);
                 obkv_writer.insert(id, val)?;
@@ -872,6 +915,15 @@ impl<'a, 'i> Transform<'a, 'i> {
                 obkv_writer.insert(id, val)?;
             }
         }
+        if !injected_vectors.is_empty() {
+            'inject_vectors: {
+                let Some(vectors_fid) = old_vectors_fid else { break 'inject_vectors };
+
+                operations.insert(vectors_fid, DelAddOperation::DeletionAndAddition);
+                obkv_writer.insert(vectors_fid, serde_json::to_vec(&injected_vectors).unwrap())?;
+            }
+        }
+
         let data = obkv_writer.into_inner()?;
         let obkv = KvReader::<FieldId>::new(&data);
 
@@ -937,6 +989,35 @@ impl<'a, 'i> Transform<'a, 'i> {
             None
         };
 
+        let readers: Result<
+            BTreeMap<&str, (Vec<arroy::Reader<arroy::distances::Angular>>, &RoaringBitmap)>,
+        > = settings_diff
+            .embedding_config_updates
+            .iter()
+            .filter_map(|(name, action)| {
+                if let EmbedderAction::WriteBackToDocuments(WriteBackToDocuments {
+                    embedder_id,
+                    user_provided,
+                }) = action
+                {
+                    let readers: Result<Vec<_>> =
+                        self.index.arroy_readers(wtxn, *embedder_id).collect();
+                    match readers {
+                        Ok(readers) => Some(Ok((name.as_str(), (readers, user_provided)))),
+                        Err(error) => Some(Err(error)),
+                    }
+                } else {
+                    None
+                }
+            })
+            .collect();
+        let readers = readers?;
+
+        let old_vectors_fid = settings_diff
+            .old
+            .fields_ids_map
+            .id(crate::vector::parsed_vectors::RESERVED_VECTORS_FIELD_NAME);
+
         // We initialize the sorter with the user indexing settings.
         let mut flattened_sorter =
             if settings_diff.reindex_searchable() || settings_diff.reindex_facets() {
@@ -963,10 +1044,50 @@ impl<'a, 'i> Transform<'a, 'i> {
                     InternalError::DatabaseMissingEntry { db_name: db_name::DOCUMENTS, key: None },
                 )?;
 
+                let injected_vectors: std::result::Result<
+                    serde_json::Map<String, serde_json::Value>,
+                    arroy::Error,
+                > = readers
+                    .iter()
+                    .filter_map(|(name, (readers, user_provided))| {
+                        if !user_provided.contains(docid) {
+                            return None;
+                        }
+                        let mut vectors = Vec::new();
+                        for reader in readers {
+                            let Some(vector) = reader.item_vector(wtxn, docid).transpose() else {
+                                break;
+                            };
+
+                            match vector {
+                                Ok(vector) => vectors.push(vector),
+                                Err(error) => return Some(Err(error)),
+                            }
+                        }
+                        if vectors.is_empty() {
+                            return None;
+                        }
+                        Some(Ok((
+                            name.to_string(),
+                            serde_json::to_value(ExplicitVectors {
+                                embeddings: Some(VectorOrArrayOfVectors::from_array_of_vectors(
+                                    vectors,
+                                )),
+                                regenerate: false,
+                            })
+                            .unwrap(),
+                        )))
+                    })
+                    .collect();
+
+                let injected_vectors = injected_vectors?;
+
                 Self::rebind_existing_document(
                     old_obkv,
                     &settings_diff,
                     &modified_faceted_fields,
+                    injected_vectors,
+                    old_vectors_fid,
                     Some(&mut original_obkv_buffer).filter(|_| original_sorter.is_some()),
                     Some(&mut flattened_obkv_buffer).filter(|_| flattened_sorter.is_some()),
                 )?;
@@ -983,6 +1104,23 @@ impl<'a, 'i> Transform<'a, 'i> {
             }
         }
 
+        let mut writers = Vec::new();
+
+        // delete all vectors from the embedders that need removal
+        for (_, (readers, _)) in readers {
+            for reader in readers {
+                let dimensions = reader.dimensions();
+                let arroy_index = reader.index();
+                drop(reader);
+                let writer = arroy::Writer::new(self.index.vector_arroy, arroy_index, dimensions);
+                writers.push(writer);
+            }
+        }
+
+        for writer in writers {
+            writer.clear(wtxn)?;
+        }
+
         let grenad_params = GrenadParameters {
             chunk_compression_type: self.indexer_settings.chunk_compression_type,
             chunk_compression_level: self.indexer_settings.chunk_compression_level,

milli/src/update/index_documents/typed_chunk.rs

@@ -20,6 +20,7 @@ use super::MergeFn;
 use crate::external_documents_ids::{DocumentOperation, DocumentOperationKind};
 use crate::facet::FacetType;
 use crate::index::db_name::DOCUMENTS;
+use crate::index::IndexEmbeddingConfig;
 use crate::proximity::MAX_DISTANCE;
 use crate::update::del_add::{deladd_serialize_add_side, DelAdd, KvReaderDelAdd};
 use crate::update::facet::FacetsUpdate;
@@ -90,6 +91,8 @@ pub(crate) enum TypedChunk {
         expected_dimension: usize,
         manual_vectors: grenad::Reader<BufReader<File>>,
         embedder_name: String,
+        add_to_user_provided: RoaringBitmap,
+        remove_from_user_provided: RoaringBitmap,
     },
     ScriptLanguageDocids(HashMap<(Script, Language), (RoaringBitmap, RoaringBitmap)>),
 }
@@ -154,8 +157,11 @@ pub(crate) fn write_typed_chunk_into_index(
             let mut docids = index.documents_ids(wtxn)?;
             let mut iter = merger.into_stream_merger_iter()?;
 
-            let embedders: BTreeSet<_> =
-                index.embedding_configs(wtxn)?.into_iter().map(|(k, _v)| k).collect();
+            let embedders: BTreeSet<_> = index
+                .embedding_configs(wtxn)?
+                .into_iter()
+                .map(|IndexEmbeddingConfig { name, .. }| name)
+                .collect();
             let mut vectors_buffer = Vec::new();
             while let Some((key, reader)) = iter.next()? {
                 let mut writer: KvWriter<_, FieldId> = KvWriter::memory();
@@ -181,7 +187,7 @@ pub(crate) fn write_typed_chunk_into_index(
                                     // if the `_vectors` field cannot be parsed as map of vectors, just write it as-is
                                     break 'vectors Some(addition);
                                 };
-                                vectors.retain_user_provided_vectors(&embedders);
+                                vectors.retain_not_embedded_vectors(&embedders);
                                 let crate::vector::parsed_vectors::ParsedVectors(vectors) = vectors;
                                 if vectors.is_empty() {
                                     // skip writing empty `_vectors` map
@@ -619,6 +625,8 @@ pub(crate) fn write_typed_chunk_into_index(
             let mut remove_vectors_builder = MergerBuilder::new(keep_first as MergeFn);
             let mut manual_vectors_builder = MergerBuilder::new(keep_first as MergeFn);
             let mut embeddings_builder = MergerBuilder::new(keep_first as MergeFn);
+            let mut add_to_user_provided = RoaringBitmap::new();
+            let mut remove_from_user_provided = RoaringBitmap::new();
             let mut params = None;
             for typed_chunk in typed_chunks {
                 let TypedChunk::VectorPoints {
@@ -627,6 +635,8 @@ pub(crate) fn write_typed_chunk_into_index(
                     embeddings,
                     expected_dimension,
                     embedder_name,
+                    add_to_user_provided: aud,
+                    remove_from_user_provided: rud,
                 } = typed_chunk
                 else {
                     unreachable!();
@@ -639,11 +649,23 @@ pub(crate) fn write_typed_chunk_into_index(
                 if let Some(embeddings) = embeddings {
                     embeddings_builder.push(embeddings.into_cursor()?);
                 }
+                add_to_user_provided |= aud;
+                remove_from_user_provided |= rud;
             }
 
             // typed chunks has always at least 1 chunk.
             let Some((expected_dimension, embedder_name)) = params else { unreachable!() };
 
+            let mut embedding_configs = index.embedding_configs(wtxn)?;
+            let index_embedder_config = embedding_configs
+                .iter_mut()
+                .find(|IndexEmbeddingConfig { name, .. }| name == &embedder_name)
+                .unwrap();
+            index_embedder_config.user_provided -= remove_from_user_provided;
+            index_embedder_config.user_provided |= add_to_user_provided;
+
+            index.put_embedding_configs(wtxn, embedding_configs)?;
+
             let embedder_index = index.embedder_category_id.get(wtxn, &embedder_name)?.ok_or(
                 InternalError::DatabaseMissingEntry { db_name: "embedder_category_id", key: None },
             )?;

milli/src/update/settings.rs

@@ -6,6 +6,7 @@ use std::sync::Arc;
 use charabia::{Normalize, Tokenizer, TokenizerBuilder};
 use deserr::{DeserializeError, Deserr};
 use itertools::{EitherOrBoth, Itertools};
+use roaring::RoaringBitmap;
 use serde::{Deserialize, Deserializer, Serialize, Serializer};
 use time::OffsetDateTime;
 
@@ -14,12 +15,18 @@ use super::index_documents::{IndexDocumentsConfig, Transform};
 use super::IndexerConfig;
 use crate::criterion::Criterion;
 use crate::error::UserError;
-use crate::index::{DEFAULT_MIN_WORD_LEN_ONE_TYPO, DEFAULT_MIN_WORD_LEN_TWO_TYPOS};
+use crate::index::{
+    IndexEmbeddingConfig, DEFAULT_MIN_WORD_LEN_ONE_TYPO, DEFAULT_MIN_WORD_LEN_TWO_TYPOS,
+};
 use crate::order_by_map::OrderByMap;
 use crate::proximity::ProximityPrecision;
 use crate::update::index_documents::IndexDocumentsMethod;
 use crate::update::{IndexDocuments, UpdateIndexingStep};
-use crate::vector::settings::{check_set, check_unset, EmbedderSource, EmbeddingSettings};
+use crate::vector::parsed_vectors::RESERVED_VECTORS_FIELD_NAME;
+use crate::vector::settings::{
+    check_set, check_unset, EmbedderAction, EmbedderSource, EmbeddingSettings, ReindexAction,
+    WriteBackToDocuments,
+};
 use crate::vector::{Embedder, EmbeddingConfig, EmbeddingConfigs};
 use crate::{FieldId, FieldsIdsMap, Index, Result};
 
@@ -490,6 +497,7 @@ impl<'a, 't, 'i> Settings<'a, 't, 'i> {
                 self.index.put_all_searchable_fields_from_fields_ids_map(
                     self.wtxn,
                     &names,
+                    &fields_ids_map.nested_ids(RESERVED_VECTORS_FIELD_NAME),
                     &fields_ids_map,
                 )?;
                 self.index.put_fields_ids_map(self.wtxn, &fields_ids_map)?;
@@ -919,92 +927,177 @@ impl<'a, 't, 'i> Settings<'a, 't, 'i> {
         Ok(changed)
     }
 
-    fn update_embedding_configs(&mut self) -> Result<bool> {
-        let update = match std::mem::take(&mut self.embedder_settings) {
-            Setting::Set(configs) => {
-                let mut changed = false;
+    fn update_embedding_configs(&mut self) -> Result<BTreeMap<String, EmbedderAction>> {
+        match std::mem::take(&mut self.embedder_settings) {
+            Setting::Set(configs) => self.update_embedding_configs_set(configs),
+            Setting::Reset => {
+                // all vectors should be written back to documents
                 let old_configs = self.index.embedding_configs(self.wtxn)?;
-                let old_configs: BTreeMap<String, Setting<EmbeddingSettings>> =
-                    old_configs.into_iter().map(|(k, v)| (k, Setting::Set(v.into()))).collect();
-
-                let mut new_configs = BTreeMap::new();
-                for joined in old_configs
+                let remove_all: Result<BTreeMap<String, EmbedderAction>> = old_configs
                     .into_iter()
-                    .merge_join_by(configs.into_iter(), |(left, _), (right, _)| left.cmp(right))
-                {
-                    match joined {
-                        // updated config
-                        EitherOrBoth::Both((name, mut old), (_, new)) => {
-                            changed |= EmbeddingSettings::apply_and_need_reindex(&mut old, new);
-                            if changed {
-                                tracing::debug!(embedder = name, "need reindex");
-                            } else {
-                                tracing::debug!(embedder = name, "skip reindex");
-                            }
-                            let new = validate_embedding_settings(old, &name)?;
-                            new_configs.insert(name, new);
-                        }
-                        // unchanged config
-                        EitherOrBoth::Left((name, setting)) => {
-                            new_configs.insert(name, setting);
-                        }
-                        // new config
-                        EitherOrBoth::Right((name, mut setting)) => {
-                            // apply the default source in case the source was not set so that it gets validated
-                            crate::vector::settings::EmbeddingSettings::apply_default_source(
-                                &mut setting,
-                            );
-                            crate::vector::settings::EmbeddingSettings::apply_default_openai_model(
-                                &mut setting,
-                            );
-                            let setting = validate_embedding_settings(setting, &name)?;
-                            changed = true;
-                            new_configs.insert(name, setting);
-                        }
-                    }
-                }
-                let new_configs: Vec<(String, EmbeddingConfig)> = new_configs
-                    .into_iter()
-                    .filter_map(|(name, setting)| match setting {
-                        Setting::Set(value) => Some((name, value.into())),
-                        Setting::Reset => None,
-                        Setting::NotSet => Some((name, EmbeddingSettings::default().into())),
+                    .map(|IndexEmbeddingConfig { name, config: _, user_provided }| -> Result<_> {
+                        let embedder_id =
+                            self.index.embedder_category_id.get(self.wtxn, &name)?.ok_or(
+                                crate::InternalError::DatabaseMissingEntry {
+                                    db_name: crate::index::db_name::VECTOR_EMBEDDER_CATEGORY_ID,
+                                    key: None,
+                                },
+                            )?;
+                        Ok((
+                            name,
+                            EmbedderAction::WriteBackToDocuments(WriteBackToDocuments {
+                                embedder_id,
+                                user_provided,
+                            }),
+                        ))
                     })
                     .collect();
 
+                let remove_all = remove_all?;
+
                 self.index.embedder_category_id.clear(self.wtxn)?;
-                for (index, (embedder_name, _)) in new_configs.iter().enumerate() {
-                    self.index.embedder_category_id.put_with_flags(
-                        self.wtxn,
-                        heed::PutFlags::APPEND,
-                        embedder_name,
-                        &index
-                            .try_into()
-                            .map_err(|_| UserError::TooManyEmbedders(new_configs.len()))?,
-                    )?;
-                }
-
-                if new_configs.is_empty() {
-                    self.index.delete_embedding_configs(self.wtxn)?;
-                } else {
-                    self.index.put_embedding_configs(self.wtxn, new_configs)?;
-                }
-                changed
-            }
-            Setting::Reset => {
                 self.index.delete_embedding_configs(self.wtxn)?;
-                true
+                Ok(remove_all)
             }
-            Setting::NotSet => false,
-        };
-
-        // if any changes force a reindexing
-        // clear the vector database.
-        if update {
-            self.index.vector_arroy.clear(self.wtxn)?;
+            Setting::NotSet => Ok(Default::default()),
         }
+    }
 
-        Ok(update)
+    fn update_embedding_configs_set(
+        &mut self,
+        configs: BTreeMap<String, Setting<EmbeddingSettings>>,
+    ) -> Result<BTreeMap<String, EmbedderAction>> {
+        use crate::vector::settings::SettingsDiff;
+
+        let old_configs = self.index.embedding_configs(self.wtxn)?;
+        let old_configs: BTreeMap<String, (EmbeddingSettings, RoaringBitmap)> = old_configs
+            .into_iter()
+            .map(|IndexEmbeddingConfig { name, config, user_provided }| {
+                (name, (config.into(), user_provided))
+            })
+            .collect();
+        let mut updated_configs = BTreeMap::new();
+        let mut embedder_actions = BTreeMap::new();
+        for joined in old_configs
+            .into_iter()
+            .merge_join_by(configs.into_iter(), |(left, _), (right, _)| left.cmp(right))
+        {
+            match joined {
+                // updated config
+                EitherOrBoth::Both((name, (old, user_provided)), (_, new)) => {
+                    let settings_diff = SettingsDiff::from_settings(old, new);
+                    match settings_diff {
+                        SettingsDiff::Remove => {
+                            tracing::debug!(
+                                embedder = name,
+                                user_provided = user_provided.len(),
+                                "removing embedder"
+                            );
+                            let embedder_id =
+                                self.index.embedder_category_id.get(self.wtxn, &name)?.ok_or(
+                                    crate::InternalError::DatabaseMissingEntry {
+                                        db_name: crate::index::db_name::VECTOR_EMBEDDER_CATEGORY_ID,
+                                        key: None,
+                                    },
+                                )?;
+                            // free id immediately
+                            self.index.embedder_category_id.delete(self.wtxn, &name)?;
+                            embedder_actions.insert(
+                                name,
+                                EmbedderAction::WriteBackToDocuments(WriteBackToDocuments {
+                                    embedder_id,
+                                    user_provided,
+                                }),
+                            );
+                        }
+                        SettingsDiff::Reindex { action, updated_settings } => {
+                            tracing::debug!(
+                                embedder = name,
+                                user_provided = user_provided.len(),
+                                ?action,
+                                "reindex embedder"
+                            );
+                            embedder_actions.insert(name.clone(), EmbedderAction::Reindex(action));
+                            let new =
+                                validate_embedding_settings(Setting::Set(updated_settings), &name)?;
+                            updated_configs.insert(name, (new, user_provided));
+                        }
+                        SettingsDiff::UpdateWithoutReindex { updated_settings } => {
+                            tracing::debug!(
+                                embedder = name,
+                                user_provided = user_provided.len(),
+                                "update without reindex embedder"
+                            );
+                            let new =
+                                validate_embedding_settings(Setting::Set(updated_settings), &name)?;
+                            updated_configs.insert(name, (new, user_provided));
+                        }
+                    }
+                }
+                // unchanged config
+                EitherOrBoth::Left((name, (setting, user_provided))) => {
+                    tracing::debug!(embedder = name, "unchanged embedder");
+                    updated_configs.insert(name, (Setting::Set(setting), user_provided));
+                }
+                // new config
+                EitherOrBoth::Right((name, mut setting)) => {
+                    tracing::debug!(embedder = name, "new embedder");
+                    // apply the default source in case the source was not set so that it gets validated
+                    crate::vector::settings::EmbeddingSettings::apply_default_source(&mut setting);
+                    crate::vector::settings::EmbeddingSettings::apply_default_openai_model(
+                        &mut setting,
+                    );
+                    let setting = validate_embedding_settings(setting, &name)?;
+                    embedder_actions
+                        .insert(name.clone(), EmbedderAction::Reindex(ReindexAction::FullReindex));
+                    updated_configs.insert(name, (setting, RoaringBitmap::new()));
+                }
+            }
+        }
+        let mut free_indices: [bool; u8::MAX as usize] = [true; u8::MAX as usize];
+        for res in self.index.embedder_category_id.iter(self.wtxn)? {
+            let (_name, id) = res?;
+            free_indices[id as usize] = false;
+        }
+        let mut free_indices = free_indices.iter_mut().enumerate();
+        let mut find_free_index =
+            move || free_indices.find(|(_, free)| **free).map(|(index, _)| index as u8);
+        for (name, action) in embedder_actions.iter() {
+            match action {
+                EmbedderAction::Reindex(ReindexAction::RegeneratePrompts) => {
+                    /* cannot be a new embedder, so has to have an id already */
+                }
+                EmbedderAction::Reindex(ReindexAction::FullReindex) => {
+                    if self.index.embedder_category_id.get(self.wtxn, name)?.is_none() {
+                        let id = find_free_index()
+                            .ok_or(UserError::TooManyEmbedders(updated_configs.len()))?;
+                        tracing::debug!(embedder = name, id, "assigning free id to new embedder");
+                        self.index.embedder_category_id.put(self.wtxn, name, &id)?;
+                    }
+                }
+                EmbedderAction::WriteBackToDocuments(_) => { /* already removed */ }
+            }
+        }
+        let updated_configs: Vec<IndexEmbeddingConfig> = updated_configs
+            .into_iter()
+            .filter_map(|(name, (config, user_provided))| match config {
+                Setting::Set(config) => {
+                    Some(IndexEmbeddingConfig { name, config: config.into(), user_provided })
+                }
+                Setting::Reset => None,
+                Setting::NotSet => Some(IndexEmbeddingConfig {
+                    name,
+                    config: EmbeddingSettings::default().into(),
+                    user_provided,
+                }),
+            })
+            .collect();
+        if updated_configs.is_empty() {
+            self.index.delete_embedding_configs(self.wtxn)?;
+        } else {
+            self.index.put_embedding_configs(self.wtxn, updated_configs)?;
+        }
+        Ok(embedder_actions)
     }
 
     fn update_search_cutoff(&mut self) -> Result<bool> {
@@ -1058,13 +1151,8 @@ impl<'a, 't, 'i> Settings<'a, 't, 'i> {
         self.update_searchable()?;
         self.update_exact_attributes()?;
         self.update_proximity_precision()?;
-        // TODO: very rough approximation of the needs for reindexing where any change will result in
-        // a full reindexing.
-        // What can be done instead:
-        // 1. Only change the distance on a distance change
-        // 2. Only change the name -> embedder mapping on a name change
-        // 3. Keep the old vectors but reattempt indexing on a prompt change: only actually changed prompt will need embedding + storage
-        let embedding_configs_updated = self.update_embedding_configs()?;
+
+        let embedding_config_updates = self.update_embedding_configs()?;
 
         let mut new_inner_settings = InnerIndexSettings::from_index(self.index, self.wtxn)?;
         new_inner_settings.recompute_facets(self.wtxn, self.index)?;
@@ -1078,7 +1166,7 @@ impl<'a, 't, 'i> Settings<'a, 't, 'i> {
             old_inner_settings,
             new_inner_settings,
             primary_key_id,
-            embedding_configs_updated,
+            embedding_config_updates,
             settings_update_only,
         );
 
@@ -1094,8 +1182,7 @@ pub struct InnerIndexSettingsDiff {
     pub(crate) old: InnerIndexSettings,
     pub(crate) new: InnerIndexSettings,
     pub(crate) primary_key_id: Option<FieldId>,
-    // TODO: compare directly the embedders.
-    pub(crate) embedding_configs_updated: bool,
+    pub(crate) embedding_config_updates: BTreeMap<String, EmbedderAction>,
     pub(crate) settings_update_only: bool,
     /// The set of only the additional searchable fields.
     /// If any other searchable field has been modified, is set to None.
@@ -1116,7 +1203,7 @@ impl InnerIndexSettingsDiff {
         old_settings: InnerIndexSettings,
         new_settings: InnerIndexSettings,
         primary_key_id: Option<FieldId>,
-        embedding_configs_updated: bool,
+        embedding_config_updates: BTreeMap<String, EmbedderAction>,
         settings_update_only: bool,
     ) -> Self {
         let only_additional_fields = match (
@@ -1153,7 +1240,7 @@ impl InnerIndexSettingsDiff {
             old: old_settings,
             new: new_settings,
             primary_key_id,
-            embedding_configs_updated,
+            embedding_config_updates,
             settings_update_only,
             only_additional_fields,
             cache_reindex_searchable_without_user_defined,
@@ -1220,7 +1307,7 @@ impl InnerIndexSettingsDiff {
     }
 
     pub fn reindex_vectors(&self) -> bool {
-        self.embedding_configs_updated
+        !self.embedding_config_updates.is_empty()
     }
 
     pub fn settings_update_only(&self) -> bool {
@@ -1252,6 +1339,8 @@ pub(crate) struct InnerIndexSettings {
     pub embedding_configs: EmbeddingConfigs,
     pub existing_fields: HashSet<String>,
     pub geo_fields_ids: Option<(FieldId, FieldId)>,
+    pub non_searchable_fields_ids: Vec<FieldId>,
+    pub non_faceted_fields_ids: Vec<FieldId>,
 }
 
 impl InnerIndexSettings {
@@ -1265,8 +1354,8 @@ impl InnerIndexSettings {
         let user_defined_searchable_fields =
             user_defined_searchable_fields.map(|sf| sf.into_iter().map(String::from).collect());
         let user_defined_faceted_fields = index.user_defined_faceted_fields(rtxn)?;
-        let searchable_fields_ids = index.searchable_fields_ids(rtxn)?;
-        let faceted_fields_ids = index.faceted_fields_ids(rtxn)?;
+        let mut searchable_fields_ids = index.searchable_fields_ids(rtxn)?;
+        let mut faceted_fields_ids = index.faceted_fields_ids(rtxn)?;
         let exact_attributes = index.exact_attributes_ids(rtxn)?;
         let proximity_precision = index.proximity_precision(rtxn)?.unwrap_or_default();
         let embedding_configs = embedders(index.embedding_configs(rtxn)?)?;
@@ -1294,6 +1383,10 @@ impl InnerIndexSettings {
             None => None,
         };
 
+        let vectors_fids = fields_ids_map.nested_ids(RESERVED_VECTORS_FIELD_NAME);
+        searchable_fields_ids.retain(|id| !vectors_fids.contains(id));
+        faceted_fields_ids.retain(|id| !vectors_fids.contains(id));
+
         Ok(Self {
             stop_words,
             allowed_separators,
@@ -1308,6 +1401,8 @@ impl InnerIndexSettings {
             embedding_configs,
             existing_fields,
             geo_fields_ids,
+            non_searchable_fields_ids: vectors_fids.clone(),
+            non_faceted_fields_ids: vectors_fids.clone(),
         })
     }
 
@@ -1315,9 +1410,10 @@ impl InnerIndexSettings {
     pub fn recompute_facets(&mut self, wtxn: &mut heed::RwTxn, index: &Index) -> Result<()> {
         let new_facets = self
             .fields_ids_map
-            .names()
-            .filter(|&field| crate::is_faceted(field, &self.user_defined_faceted_fields))
-            .map(|field| field.to_string())
+            .iter()
+            .filter(|(fid, _field)| !self.non_faceted_fields_ids.contains(fid))
+            .filter(|(_fid, field)| crate::is_faceted(field, &self.user_defined_faceted_fields))
+            .map(|(_fid, field)| field.to_string())
             .collect();
         index.put_faceted_fields(wtxn, &new_facets)?;
 
@@ -1337,6 +1433,7 @@ impl InnerIndexSettings {
             index.put_all_searchable_fields_from_fields_ids_map(
                 wtxn,
                 &searchable_fields,
+                &self.non_searchable_fields_ids,
                 &self.fields_ids_map,
             )?;
         }
@@ -1347,19 +1444,25 @@ impl InnerIndexSettings {
     }
 }
 
-fn embedders(embedding_configs: Vec<(String, EmbeddingConfig)>) -> Result<EmbeddingConfigs> {
+fn embedders(embedding_configs: Vec<IndexEmbeddingConfig>) -> Result<EmbeddingConfigs> {
     let res: Result<_> = embedding_configs
         .into_iter()
-        .map(|(name, EmbeddingConfig { embedder_options, prompt })| {
-            let prompt = Arc::new(prompt.try_into().map_err(crate::Error::from)?);
+        .map(
+            |IndexEmbeddingConfig {
+                 name,
+                 config: EmbeddingConfig { embedder_options, prompt },
+                 ..
+             }| {
+                let prompt = Arc::new(prompt.try_into().map_err(crate::Error::from)?);
 
-            let embedder = Arc::new(
-                Embedder::new(embedder_options.clone())
-                    .map_err(crate::vector::Error::from)
-                    .map_err(crate::Error::from)?,
-            );
-            Ok((name, (embedder, prompt)))
-        })
+                let embedder = Arc::new(
+                    Embedder::new(embedder_options.clone())
+                        .map_err(crate::vector::Error::from)
+                        .map_err(crate::Error::from)?,
+                );
+                Ok((name, (embedder, prompt)))
+            },
+        )
         .collect();
     res.map(EmbeddingConfigs::new)
 }

milli/src/vector/mod.rs

@@ -152,6 +152,10 @@ impl EmbeddingConfigs {
         &self.0
     }
 
+    pub fn into_inner(self) -> HashMap<String, (Arc<Embedder>, Arc<Prompt>)> {
+        self.0
+    }
+
     /// Get the name of the default embedder configuration.
     ///
     /// The default embedder is determined as follows:

milli/src/vector/parsed_vectors.rs

@@ -4,8 +4,9 @@ use obkv::KvReader;
 use serde_json::{from_slice, Value};
 
 use super::Embedding;
+use crate::index::IndexEmbeddingConfig;
 use crate::update::del_add::{DelAdd, KvReaderDelAdd};
-use crate::{FieldId, InternalError, UserError};
+use crate::{DocumentId, FieldId, InternalError, UserError};
 
 pub const RESERVED_VECTORS_FIELD_NAME: &str = "_vectors";
 
@@ -17,11 +18,20 @@ pub enum Vectors {
 }
 
 impl Vectors {
-    pub fn into_array_of_vectors(self) -> Vec<Embedding> {
+    pub fn must_regenerate(&self) -> bool {
         match self {
-            Vectors::ImplicitlyUserProvided(embeddings)
-            | Vectors::Explicit(ExplicitVectors { embeddings, user_provided: _ }) => {
-                embeddings.into_array_of_vectors().unwrap_or_default()
+            Vectors::ImplicitlyUserProvided(_) => false,
+            Vectors::Explicit(ExplicitVectors { regenerate, .. }) => *regenerate,
+        }
+    }
+
+    pub fn into_array_of_vectors(self) -> Option<Vec<Embedding>> {
+        match self {
+            Vectors::ImplicitlyUserProvided(embeddings) => {
+                Some(embeddings.into_array_of_vectors().unwrap_or_default())
+            }
+            Vectors::Explicit(ExplicitVectors { embeddings, regenerate: _ }) => {
+                embeddings.map(|embeddings| embeddings.into_array_of_vectors().unwrap_or_default())
             }
         }
     }
@@ -30,22 +40,46 @@ impl Vectors {
 #[derive(serde::Serialize, serde::Deserialize, Debug)]
 #[serde(rename_all = "camelCase")]
 pub struct ExplicitVectors {
-    pub embeddings: VectorOrArrayOfVectors,
-    pub user_provided: bool,
+    pub embeddings: Option<VectorOrArrayOfVectors>,
+    pub regenerate: bool,
+}
+
+pub enum VectorState {
+    Inline(Vectors),
+    Manual,
+    Generated,
+}
+
+impl VectorState {
+    pub fn must_regenerate(&self) -> bool {
+        match self {
+            VectorState::Inline(vectors) => vectors.must_regenerate(),
+            VectorState::Manual => false,
+            VectorState::Generated => true,
+        }
+    }
+}
+
+pub enum VectorsState {
+    NoVectorsFid,
+    NoVectorsFieldInDocument,
+    Vectors(BTreeMap<String, Vectors>),
 }
 
 pub struct ParsedVectorsDiff {
-    pub old: Option<BTreeMap<String, Vectors>>,
-    pub new: Option<BTreeMap<String, Vectors>>,
+    old: BTreeMap<String, VectorState>,
+    new: VectorsState,
 }
 
 impl ParsedVectorsDiff {
     pub fn new(
+        docid: DocumentId,
+        embedders_configs: &[IndexEmbeddingConfig],
         documents_diff: KvReader<'_, FieldId>,
         old_vectors_fid: Option<FieldId>,
         new_vectors_fid: Option<FieldId>,
     ) -> Result<Self, Error> {
-        let old = match old_vectors_fid
+        let mut old = match old_vectors_fid
             .and_then(|vectors_fid| documents_diff.get(vectors_fid))
             .map(KvReaderDelAdd::new)
             .map(|obkv| to_vector_map(obkv, DelAdd::Deletion))
@@ -61,19 +95,54 @@ impl ParsedVectorsDiff {
                 return Err(error);
             }
         }
-        .flatten();
-        let new = new_vectors_fid
-            .and_then(|vectors_fid| documents_diff.get(vectors_fid))
-            .map(KvReaderDelAdd::new)
-            .map(|obkv| to_vector_map(obkv, DelAdd::Addition))
-            .transpose()?
-            .flatten();
+        .flatten().map_or(BTreeMap::default(), |del| del.into_iter().map(|(name, vec)| (name, VectorState::Inline(vec))).collect());
+        for embedding_config in embedders_configs {
+            if embedding_config.user_provided.contains(docid) {
+                old.entry(embedding_config.name.to_string()).or_insert(VectorState::Manual);
+            }
+        }
+
+        let new = 'new: {
+            let Some(new_vectors_fid) = new_vectors_fid else {
+                break 'new VectorsState::NoVectorsFid;
+            };
+            let Some(bytes) = documents_diff.get(new_vectors_fid) else {
+                break 'new VectorsState::NoVectorsFieldInDocument;
+            };
+            let obkv = KvReaderDelAdd::new(bytes);
+            match to_vector_map(obkv, DelAdd::Addition)? {
+                Some(new) => VectorsState::Vectors(new),
+                None => VectorsState::NoVectorsFieldInDocument,
+            }
+        };
+
         Ok(Self { old, new })
     }
 
-    pub fn remove(&mut self, embedder_name: &str) -> (Option<Vectors>, Option<Vectors>) {
-        let old = self.old.as_mut().and_then(|old| old.remove(embedder_name));
-        let new = self.new.as_mut().and_then(|new| new.remove(embedder_name));
+    pub fn remove(&mut self, embedder_name: &str) -> (VectorState, VectorState) {
+        let old = self.old.remove(embedder_name).unwrap_or(VectorState::Generated);
+        let state_from_old = match old {
+            // assume a userProvided is still userProvided
+            VectorState::Manual => VectorState::Manual,
+            // generated is still generated
+            VectorState::Generated => VectorState::Generated,
+            // weird case that shouldn't happen were the previous docs version is inline,
+            // but it was removed in the new version
+            // Since it is not in the new version, we switch to generated
+            VectorState::Inline(_) => VectorState::Generated,
+        };
+        let new = match &mut self.new {
+            VectorsState::Vectors(new) => {
+                new.remove(embedder_name).map(VectorState::Inline).unwrap_or(state_from_old)
+            }
+            _ =>
+            // if no `_vectors` field is present in the new document,
+            // the state depends on the previous version of the document
+            {
+                state_from_old
+            }
+        };
+
         (old, new)
     }
 }
@@ -89,15 +158,8 @@ impl ParsedVectors {
         Ok(ParsedVectors(value))
     }
 
-    pub fn retain_user_provided_vectors(&mut self, embedders: &BTreeSet<String>) {
-        self.0.retain(|k, v| match v {
-            Vectors::ImplicitlyUserProvided(_) => true,
-            Vectors::Explicit(ExplicitVectors { embeddings: _, user_provided }) => {
-                *user_provided
-                // if the embedder is not in the config, then never touch it
-                || !embedders.contains(k)
-            }
-        });
+    pub fn retain_not_embedded_vectors(&mut self, embedders: &BTreeSet<String>) {
+        self.0.retain(|k, _v| !embedders.contains(k))
     }
 }
 
@@ -150,6 +212,22 @@ impl VectorOrArrayOfVectors {
     pub fn from_array_of_vectors(array_of_vec: Vec<Embedding>) -> Self {
         Self { inner: Some(either::Either::Left(array_of_vec)) }
     }
+
+    pub fn from_vector(vec: Embedding) -> Self {
+        Self { inner: Some(either::Either::Right(vec)) }
+    }
+}
+
+impl From<Embedding> for VectorOrArrayOfVectors {
+    fn from(vec: Embedding) -> Self {
+        Self::from_vector(vec)
+    }
+}
+
+impl From<Vec<Embedding>> for VectorOrArrayOfVectors {
+    fn from(vec: Vec<Embedding>) -> Self {
+        Self::from_array_of_vectors(vec)
+    }
 }
 
 #[cfg(test)]

milli/src/vector/settings.rs

@@ -1,4 +1,5 @@
 use deserr::Deserr;
+use roaring::RoaringBitmap;
 use serde::{Deserialize, Serialize};
 
 use super::rest::InputType;
@@ -72,6 +73,238 @@ pub fn check_unset<T>(
     }
 }
 
+/// Indicates what action should take place during a reindexing operation for an embedder
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
+pub enum ReindexAction {
+    /// An indexing operation should take place for this embedder, keeping existing vectors
+    /// and checking whether the document template changed or not
+    RegeneratePrompts,
+    /// An indexing operation should take place for all documents for this embedder, removing existing vectors
+    /// (except userProvided ones)
+    FullReindex,
+}
+
+pub enum SettingsDiff {
+    Remove,
+    Reindex { action: ReindexAction, updated_settings: EmbeddingSettings },
+    UpdateWithoutReindex { updated_settings: EmbeddingSettings },
+}
+
+pub enum EmbedderAction {
+    WriteBackToDocuments(WriteBackToDocuments),
+    Reindex(ReindexAction),
+}
+
+pub struct WriteBackToDocuments {
+    pub embedder_id: u8,
+    pub user_provided: RoaringBitmap,
+}
+
+impl SettingsDiff {
+    pub fn from_settings(old: EmbeddingSettings, new: Setting<EmbeddingSettings>) -> Self {
+        match new {
+            Setting::Set(new) => {
+                let EmbeddingSettings {
+                    mut source,
+                    mut model,
+                    mut revision,
+                    mut api_key,
+                    mut dimensions,
+                    mut document_template,
+                    mut url,
+                    mut query,
+                    mut input_field,
+                    mut path_to_embeddings,
+                    mut embedding_object,
+                    mut input_type,
+                    mut distribution,
+                } = old;
+
+                let EmbeddingSettings {
+                    source: new_source,
+                    model: new_model,
+                    revision: new_revision,
+                    api_key: new_api_key,
+                    dimensions: new_dimensions,
+                    document_template: new_document_template,
+                    url: new_url,
+                    query: new_query,
+                    input_field: new_input_field,
+                    path_to_embeddings: new_path_to_embeddings,
+                    embedding_object: new_embedding_object,
+                    input_type: new_input_type,
+                    distribution: new_distribution,
+                } = new;
+
+                let mut reindex_action = None;
+
+                // **Warning**: do not use short-circuiting || here, we want all these operations applied
+                if source.apply(new_source) {
+                    ReindexAction::push_action(&mut reindex_action, ReindexAction::FullReindex);
+                    // when the source changes, we need to reapply the default settings for the new source
+                    apply_default_for_source(
+                        &source,
+                        &mut model,
+                        &mut revision,
+                        &mut dimensions,
+                        &mut url,
+                        &mut query,
+                        &mut input_field,
+                        &mut path_to_embeddings,
+                        &mut embedding_object,
+                        &mut input_type,
+                        &mut document_template,
+                    )
+                }
+                if model.apply(new_model) {
+                    ReindexAction::push_action(&mut reindex_action, ReindexAction::FullReindex);
+                }
+                if revision.apply(new_revision) {
+                    ReindexAction::push_action(&mut reindex_action, ReindexAction::FullReindex);
+                }
+                if dimensions.apply(new_dimensions) {
+                    ReindexAction::push_action(&mut reindex_action, ReindexAction::FullReindex);
+                }
+                if url.apply(new_url) {
+                    ReindexAction::push_action(&mut reindex_action, ReindexAction::FullReindex);
+                }
+                if query.apply(new_query) {
+                    ReindexAction::push_action(&mut reindex_action, ReindexAction::FullReindex);
+                }
+                if input_field.apply(new_input_field) {
+                    ReindexAction::push_action(&mut reindex_action, ReindexAction::FullReindex);
+                }
+                if path_to_embeddings.apply(new_path_to_embeddings) {
+                    ReindexAction::push_action(&mut reindex_action, ReindexAction::FullReindex);
+                }
+                if embedding_object.apply(new_embedding_object) {
+                    ReindexAction::push_action(&mut reindex_action, ReindexAction::FullReindex);
+                }
+                if input_type.apply(new_input_type) {
+                    ReindexAction::push_action(&mut reindex_action, ReindexAction::FullReindex);
+                }
+                if document_template.apply(new_document_template) {
+                    ReindexAction::push_action(
+                        &mut reindex_action,
+                        ReindexAction::RegeneratePrompts,
+                    );
+                }
+
+                distribution.apply(new_distribution);
+                api_key.apply(new_api_key);
+
+                let updated_settings = EmbeddingSettings {
+                    source,
+                    model,
+                    revision,
+                    api_key,
+                    dimensions,
+                    document_template,
+                    url,
+                    query,
+                    input_field,
+                    path_to_embeddings,
+                    embedding_object,
+                    input_type,
+                    distribution,
+                };
+
+                match reindex_action {
+                    Some(action) => Self::Reindex { action, updated_settings },
+                    None => Self::UpdateWithoutReindex { updated_settings },
+                }
+            }
+            Setting::Reset => Self::Remove,
+            Setting::NotSet => Self::UpdateWithoutReindex { updated_settings: old },
+        }
+    }
+}
+
+impl ReindexAction {
+    fn push_action(this: &mut Option<Self>, other: Self) {
+        *this = match (*this, other) {
+            (_, ReindexAction::FullReindex) => Some(ReindexAction::FullReindex),
+            (Some(ReindexAction::FullReindex), _) => Some(ReindexAction::FullReindex),
+            (_, ReindexAction::RegeneratePrompts) => Some(ReindexAction::RegeneratePrompts),
+        }
+    }
+}
+
+#[allow(clippy::too_many_arguments)] // private function
+fn apply_default_for_source(
+    source: &Setting<EmbedderSource>,
+    model: &mut Setting<String>,
+    revision: &mut Setting<String>,
+    dimensions: &mut Setting<usize>,
+    url: &mut Setting<String>,
+    query: &mut Setting<serde_json::Value>,
+    input_field: &mut Setting<Vec<String>>,
+    path_to_embeddings: &mut Setting<Vec<String>>,
+    embedding_object: &mut Setting<Vec<String>>,
+    input_type: &mut Setting<InputType>,
+    document_template: &mut Setting<String>,
+) {
+    match source {
+        Setting::Set(EmbedderSource::HuggingFace) => {
+            *model = Setting::Reset;
+            *revision = Setting::Reset;
+            *dimensions = Setting::NotSet;
+            *url = Setting::NotSet;
+            *query = Setting::NotSet;
+            *input_field = Setting::NotSet;
+            *path_to_embeddings = Setting::NotSet;
+            *embedding_object = Setting::NotSet;
+            *input_type = Setting::NotSet;
+        }
+        Setting::Set(EmbedderSource::Ollama) => {
+            *model = Setting::Reset;
+            *revision = Setting::NotSet;
+            *dimensions = Setting::Reset;
+            *url = Setting::NotSet;
+            *query = Setting::NotSet;
+            *input_field = Setting::NotSet;
+            *path_to_embeddings = Setting::NotSet;
+            *embedding_object = Setting::NotSet;
+            *input_type = Setting::NotSet;
+        }
+        Setting::Set(EmbedderSource::OpenAi) | Setting::Reset => {
+            *model = Setting::Reset;
+            *revision = Setting::NotSet;
+            *dimensions = Setting::NotSet;
+            *url = Setting::NotSet;
+            *query = Setting::NotSet;
+            *input_field = Setting::NotSet;
+            *path_to_embeddings = Setting::NotSet;
+            *embedding_object = Setting::NotSet;
+            *input_type = Setting::NotSet;
+        }
+        Setting::Set(EmbedderSource::Rest) => {
+            *model = Setting::NotSet;
+            *revision = Setting::NotSet;
+            *dimensions = Setting::Reset;
+            *url = Setting::Reset;
+            *query = Setting::Reset;
+            *input_field = Setting::Reset;
+            *path_to_embeddings = Setting::Reset;
+            *embedding_object = Setting::Reset;
+            *input_type = Setting::Reset;
+        }
+        Setting::Set(EmbedderSource::UserProvided) => {
+            *model = Setting::NotSet;
+            *revision = Setting::NotSet;
+            *dimensions = Setting::Reset;
+            *url = Setting::NotSet;
+            *query = Setting::NotSet;
+            *input_field = Setting::NotSet;
+            *path_to_embeddings = Setting::NotSet;
+            *embedding_object = Setting::NotSet;
+            *input_type = Setting::NotSet;
+            *document_template = Setting::NotSet;
+        }
+        Setting::NotSet => {}
+    }
+}
+
 pub fn check_set<T>(
     key: &Setting<T>,
     field: &'static str,
@@ -210,66 +443,6 @@ impl EmbeddingSettings {
             *model = Setting::Set(openai::EmbeddingModel::default().name().to_owned())
         }
     }
-
-    pub(crate) fn apply_and_need_reindex(
-        old: &mut Setting<EmbeddingSettings>,
-        new: Setting<EmbeddingSettings>,
-    ) -> bool {
-        match (old, new) {
-            (
-                Setting::Set(EmbeddingSettings {
-                    source: old_source,
-                    model: old_model,
-                    revision: old_revision,
-                    api_key: old_api_key,
-                    dimensions: old_dimensions,
-                    document_template: old_document_template,
-                    url: old_url,
-                    query: old_query,
-                    input_field: old_input_field,
-                    path_to_embeddings: old_path_to_embeddings,
-                    embedding_object: old_embedding_object,
-                    input_type: old_input_type,
-                    distribution: old_distribution,
-                }),
-                Setting::Set(EmbeddingSettings {
-                    source: new_source,
-                    model: new_model,
-                    revision: new_revision,
-                    api_key: new_api_key,
-                    dimensions: new_dimensions,
-                    document_template: new_document_template,
-                    url: new_url,
-                    query: new_query,
-                    input_field: new_input_field,
-                    path_to_embeddings: new_path_to_embeddings,
-                    embedding_object: new_embedding_object,
-                    input_type: new_input_type,
-                    distribution: new_distribution,
-                }),
-            ) => {
-                let mut needs_reindex = false;
-
-                needs_reindex |= old_source.apply(new_source);
-                needs_reindex |= old_model.apply(new_model);
-                needs_reindex |= old_revision.apply(new_revision);
-                needs_reindex |= old_dimensions.apply(new_dimensions);
-                needs_reindex |= old_document_template.apply(new_document_template);
-                needs_reindex |= old_url.apply(new_url);
-                needs_reindex |= old_query.apply(new_query);
-                needs_reindex |= old_input_field.apply(new_input_field);
-                needs_reindex |= old_path_to_embeddings.apply(new_path_to_embeddings);
-                needs_reindex |= old_embedding_object.apply(new_embedding_object);
-                needs_reindex |= old_input_type.apply(new_input_type);
-
-                old_distribution.apply(new_distribution);
-                old_api_key.apply(new_api_key);
-                needs_reindex
-            }
-            (Setting::Reset, Setting::Reset) | (_, Setting::NotSet) => false,
-            _ => true,
-        }
-    }
 }
 
 #[derive(Debug, Clone, Copy, Default, Serialize, Deserialize, PartialEq, Eq, Deserr)]

xtask/Cargo.toml

@@ -21,7 +21,7 @@ reqwest = { version = "0.11.23", features = [
     "stream",
     "json",
     "rustls-tls",
-], default_features = false }
+], default-features = false }
 serde = { version = "1.0.195", features = ["derive"] }
 serde_json = "1.0.111"
 sha2 = "0.10.8"