with binary quantization

Cargo.lock

@@ -398,6 +398,24 @@ dependencies = [
  "thiserror",
 ]
 
+[[package]]
+name = "arroy"
+version = "0.4.0"
+source = "git+https://github.com/meilisearch/arroy?branch=binary-quantization#facc8575222d3f5da5b9a94288e44896911e701f"
+dependencies = [
+ "bytemuck",
+ "byteorder",
+ "heed",
+ "log",
+ "memmap2 0.9.4",
+ "ordered-float",
+ "rand",
+ "rayon",
+ "roaring",
+ "tempfile",
+ "thiserror",
+]
+
 [[package]]
 name = "assert-json-diff"
 version = "2.0.2"
@@ -2455,7 +2473,7 @@ name = "index-scheduler"
 version = "1.9.0"
 dependencies = [
  "anyhow",
- "arroy",
+ "arroy 0.4.0 (registry+https://github.com/rust-lang/crates.io-index)",
  "big_s",
  "bincode",
  "crossbeam",
@@ -3448,7 +3466,7 @@ dependencies = [
 name = "milli"
 version = "1.9.0"
 dependencies = [
- "arroy",
+ "arroy 0.4.0 (git+https://github.com/meilisearch/arroy?branch=binary-quantization)",
  "big_s",
  "bimap",
  "bincode",

milli/Cargo.toml

@@ -79,7 +79,7 @@ hf-hub = { git = "https://github.com/dureuill/hf-hub.git", branch = "rust_tls",
 ] }
 tiktoken-rs = "0.5.8"
 liquid = "0.26.4"
-arroy = "0.4.0"
+arroy = { git = "https://github.com/meilisearch/arroy", branch = "binary-quantization" }
 rand = "0.8.5"
 tracing = "0.1.40"
 ureq = { version = "2.9.7", features = ["json"] }

milli/src/index.rs

@@ -169,7 +169,7 @@ pub struct Index {
     /// Maps an embedder name to its id in the arroy store.
     pub embedder_category_id: Database<Str, U8>,
     /// Vector store based on arroy™.
-    pub vector_arroy: arroy::Database<arroy::distances::Angular>,
+    pub vector_arroy: arroy::Database<arroy::distances::BinaryQuantizedEuclidean>,
 
     /// Maps the document id to the document as an obkv store.
     pub(crate) documents: Database<BEU32, ObkvCodec>,
@@ -1605,7 +1605,8 @@ impl Index {
         &'a self,
         rtxn: &'a RoTxn<'a>,
         embedder_id: u8,
-    ) -> impl Iterator<Item = Result<arroy::Reader<arroy::distances::Angular>>> + 'a {
+    ) -> impl Iterator<Item = Result<arroy::Reader<arroy::distances::BinaryQuantizedEuclidean>>> + 'a
+    {
         crate::vector::arroy_db_range_for_embedder(embedder_id).map_while(move |k| {
             arroy::Reader::open(rtxn, k, self.vector_arroy)
                 .map(Some)

milli/src/update/index_documents/transform.rs

@@ -990,7 +990,10 @@ impl<'a, 'i> Transform<'a, 'i> {
         };
 
         let readers: Result<
-            BTreeMap<&str, (Vec<arroy::Reader<arroy::distances::Angular>>, &RoaringBitmap)>,
+            BTreeMap<
+                &str,
+                (Vec<arroy::Reader<arroy::distances::BinaryQuantizedEuclidean>>, &RoaringBitmap),
+            >,
         > = settings_diff
             .embedding_config_updates
             .iter()