diff --git a/crates/milli/src/vector/composite.rs b/crates/milli/src/vector/composite.rs
new file mode 100644
index 000000000..d174232bf
--- /dev/null
+++ b/crates/milli/src/vector/composite.rs
@@ -0,0 +1,280 @@
+use std::time::Instant;
+
+use arroy::Distance;
+
+use super::error::CompositeEmbedderContainsHuggingFace;
+use super::{
+    hf, manual, ollama, openai, rest, DistributionShift, EmbedError, Embedding, NewEmbedderError,
+};
+use crate::ThreadPoolNoAbort;
+
+#[derive(Debug)]
+pub enum SubEmbedder {
+    /// An embedder based on running local models, fetched from the Hugging Face Hub.
+    HuggingFace(hf::Embedder),
+    /// An embedder based on making embedding queries against the OpenAI API.
+    OpenAi(openai::Embedder),
+    /// An embedder based on the user providing the embeddings in the documents and queries.
+    UserProvided(manual::Embedder),
+    /// An embedder based on making embedding queries against an <https://ollama.com> embedding server.
+    Ollama(ollama::Embedder),
+    /// An embedder based on making embedding queries against a generic JSON/REST embedding server.
+    Rest(rest::Embedder),
+}
+
+#[derive(Debug, Clone, Hash, PartialEq, Eq, serde::Deserialize, serde::Serialize)]
+pub enum SubEmbedderOptions {
+    HuggingFace(hf::EmbedderOptions),
+    OpenAi(openai::EmbedderOptions),
+    Ollama(ollama::EmbedderOptions),
+    UserProvided(manual::EmbedderOptions),
+    Rest(rest::EmbedderOptions),
+}
+
+impl SubEmbedderOptions {
+    pub fn distribution(&self) -> Option<DistributionShift> {
+        match self {
+            SubEmbedderOptions::HuggingFace(embedder_options) => embedder_options.distribution,
+            SubEmbedderOptions::OpenAi(embedder_options) => embedder_options.distribution,
+            SubEmbedderOptions::Ollama(embedder_options) => embedder_options.distribution,
+            SubEmbedderOptions::UserProvided(embedder_options) => embedder_options.distribution,
+            SubEmbedderOptions::Rest(embedder_options) => embedder_options.distribution,
+        }
+    }
+}
+
+#[derive(Debug)]
+pub struct Embedder {
+    pub(super) search: SubEmbedder,
+    pub(super) index: SubEmbedder,
+}
+
+#[derive(Debug, Clone, Hash, PartialEq, Eq, serde::Deserialize, serde::Serialize)]
+pub struct EmbedderOptions {
+    pub search: SubEmbedderOptions,
+    pub index: SubEmbedderOptions,
+}
+
+impl Embedder {
+    pub fn new(
+        EmbedderOptions { search, index }: EmbedderOptions,
+    ) -> Result<Self, NewEmbedderError> {
+        let search = SubEmbedder::new(search)?;
+        let index = SubEmbedder::new(index)?;
+
+        // check dimensions
+        if search.dimensions() != index.dimensions() {
+            return Err(NewEmbedderError::composite_dimensions_mismatch(
+                search.dimensions(),
+                index.dimensions(),
+            ));
+        }
+        // check similarity
+        let search_embeddings = search
+            .embed(
+                vec![
+                    "test".into(),
+                    "a brave dog".into(),
+                    "This is a sample text. It is meant to compare similarity.".into(),
+                ],
+                None,
+            )
+            .map_err(|error| NewEmbedderError::composite_test_embedding_failed(error, "search"))?;
+
+        let index_embeddings = index
+            .embed(
+                vec![
+                    "test".into(),
+                    "a brave dog".into(),
+                    "This is a sample text. It is meant to compare similarity.".into(),
+                ],
+                None,
+            )
+            .map_err(|error| {
+                NewEmbedderError::composite_test_embedding_failed(error, "indexing")
+            })?;
+
+        let hint = configuration_hint(&search, &index);
+
+        check_similarity(search_embeddings, index_embeddings, hint)?;
+
+        Ok(Self { search, index })
+    }
+
+    /// Indicates the dimensions of a single embedding produced by the embedder.
+    pub fn dimensions(&self) -> usize {
+        // can use the dimensions of any embedder since they should match
+        self.index.dimensions()
+    }
+
+    /// An optional distribution used to apply an affine transformation to the similarity score of a document.
+    pub fn distribution(&self) -> Option<DistributionShift> {
+        // 3 cases here:
+        // 1. distribution provided by user => use that one, which was stored in search
+        // 2. no user-provided distribution, distribution in search embedder => use that one
+        // 2. no user-provided distribution, no distribution in search embedder => use the distribution in indexing embedder
+        self.search.distribution().or_else(|| self.index.distribution())
+    }
+}
+
+impl SubEmbedder {
+    pub fn new(options: SubEmbedderOptions) -> std::result::Result<Self, NewEmbedderError> {
+        Ok(match options {
+            SubEmbedderOptions::HuggingFace(options) => {
+                Self::HuggingFace(hf::Embedder::new(options)?)
+            }
+            SubEmbedderOptions::OpenAi(options) => Self::OpenAi(openai::Embedder::new(options)?),
+            SubEmbedderOptions::Ollama(options) => Self::Ollama(ollama::Embedder::new(options)?),
+            SubEmbedderOptions::UserProvided(options) => {
+                Self::UserProvided(manual::Embedder::new(options))
+            }
+            SubEmbedderOptions::Rest(options) => {
+                Self::Rest(rest::Embedder::new(options, rest::ConfigurationSource::User)?)
+            }
+        })
+    }
+
+    pub fn embed(
+        &self,
+        texts: Vec<String>,
+        deadline: Option<Instant>,
+    ) -> std::result::Result<Vec<Embedding>, EmbedError> {
+        match self {
+            SubEmbedder::HuggingFace(embedder) => embedder.embed(texts),
+            SubEmbedder::OpenAi(embedder) => embedder.embed(&texts, deadline),
+            SubEmbedder::Ollama(embedder) => embedder.embed(&texts, deadline),
+            SubEmbedder::UserProvided(embedder) => embedder.embed(&texts),
+            SubEmbedder::Rest(embedder) => embedder.embed(texts, deadline),
+        }
+    }
+
+    /// Embed multiple chunks of texts.
+    ///
+    /// Each chunk is composed of one or multiple texts.
+    pub fn embed_index(
+        &self,
+        text_chunks: Vec<Vec<String>>,
+        threads: &ThreadPoolNoAbort,
+    ) -> std::result::Result<Vec<Vec<Embedding>>, EmbedError> {
+        match self {
+            SubEmbedder::HuggingFace(embedder) => embedder.embed_index(text_chunks),
+            SubEmbedder::OpenAi(embedder) => embedder.embed_index(text_chunks, threads),
+            SubEmbedder::Ollama(embedder) => embedder.embed_index(text_chunks, threads),
+            SubEmbedder::UserProvided(embedder) => embedder.embed_index(text_chunks),
+            SubEmbedder::Rest(embedder) => embedder.embed_index(text_chunks, threads),
+        }
+    }
+
+    /// Non-owning variant of [`Self::embed_index`].
+    pub fn embed_index_ref(
+        &self,
+        texts: &[&str],
+        threads: &ThreadPoolNoAbort,
+    ) -> std::result::Result<Vec<Embedding>, EmbedError> {
+        match self {
+            SubEmbedder::HuggingFace(embedder) => embedder.embed_index_ref(texts),
+            SubEmbedder::OpenAi(embedder) => embedder.embed_index_ref(texts, threads),
+            SubEmbedder::Ollama(embedder) => embedder.embed_index_ref(texts, threads),
+            SubEmbedder::UserProvided(embedder) => embedder.embed_index_ref(texts),
+            SubEmbedder::Rest(embedder) => embedder.embed_index_ref(texts, threads),
+        }
+    }
+
+    /// Indicates the preferred number of chunks to pass to [`Self::embed_chunks`]
+    pub fn chunk_count_hint(&self) -> usize {
+        match self {
+            SubEmbedder::HuggingFace(embedder) => embedder.chunk_count_hint(),
+            SubEmbedder::OpenAi(embedder) => embedder.chunk_count_hint(),
+            SubEmbedder::Ollama(embedder) => embedder.chunk_count_hint(),
+            SubEmbedder::UserProvided(_) => 100,
+            SubEmbedder::Rest(embedder) => embedder.chunk_count_hint(),
+        }
+    }
+
+    /// Indicates the preferred number of texts in a single chunk passed to [`Self::embed`]
+    pub fn prompt_count_in_chunk_hint(&self) -> usize {
+        match self {
+            SubEmbedder::HuggingFace(embedder) => embedder.prompt_count_in_chunk_hint(),
+            SubEmbedder::OpenAi(embedder) => embedder.prompt_count_in_chunk_hint(),
+            SubEmbedder::Ollama(embedder) => embedder.prompt_count_in_chunk_hint(),
+            SubEmbedder::UserProvided(_) => 1,
+            SubEmbedder::Rest(embedder) => embedder.prompt_count_in_chunk_hint(),
+        }
+    }
+
+    pub fn uses_document_template(&self) -> bool {
+        match self {
+            SubEmbedder::HuggingFace(_)
+            | SubEmbedder::OpenAi(_)
+            | SubEmbedder::Ollama(_)
+            | SubEmbedder::Rest(_) => true,
+            SubEmbedder::UserProvided(_) => false,
+        }
+    }
+
+    /// Indicates the dimensions of a single embedding produced by the embedder.
+    pub fn dimensions(&self) -> usize {
+        match self {
+            SubEmbedder::HuggingFace(embedder) => embedder.dimensions(),
+            SubEmbedder::OpenAi(embedder) => embedder.dimensions(),
+            SubEmbedder::Ollama(embedder) => embedder.dimensions(),
+            SubEmbedder::UserProvided(embedder) => embedder.dimensions(),
+            SubEmbedder::Rest(embedder) => embedder.dimensions(),
+        }
+    }
+
+    /// An optional distribution used to apply an affine transformation to the similarity score of a document.
+    pub fn distribution(&self) -> Option<DistributionShift> {
+        match self {
+            SubEmbedder::HuggingFace(embedder) => embedder.distribution(),
+            SubEmbedder::OpenAi(embedder) => embedder.distribution(),
+            SubEmbedder::Ollama(embedder) => embedder.distribution(),
+            SubEmbedder::UserProvided(embedder) => embedder.distribution(),
+            SubEmbedder::Rest(embedder) => embedder.distribution(),
+        }
+    }
+}
+
+fn check_similarity(
+    left: Vec<Embedding>,
+    right: Vec<Embedding>,
+    hint: CompositeEmbedderContainsHuggingFace,
+) -> Result<(), NewEmbedderError> {
+    if left.len() != right.len() {
+        return Err(NewEmbedderError::composite_embedding_count_mismatch(left.len(), right.len()));
+    }
+
+    for (left, right) in left.into_iter().zip(right) {
+        let left = arroy::internals::UnalignedVector::from_slice(&left);
+        let right = arroy::internals::UnalignedVector::from_slice(&right);
+        let left = arroy::internals::Leaf {
+            header: arroy::distances::Cosine::new_header(&left),
+            vector: left,
+        };
+        let right = arroy::internals::Leaf {
+            header: arroy::distances::Cosine::new_header(&right),
+            vector: right,
+        };
+
+        let distance = arroy::distances::Cosine::built_distance(&left, &right);
+
+        if distance > super::MAX_COMPOSITE_DISTANCE {
+            return Err(NewEmbedderError::composite_embedding_value_mismatch(distance, hint));
+        }
+    }
+    Ok(())
+}
+
+fn configuration_hint(
+    search: &SubEmbedder,
+    index: &SubEmbedder,
+) -> CompositeEmbedderContainsHuggingFace {
+    match (search, index) {
+        (SubEmbedder::HuggingFace(_), SubEmbedder::HuggingFace(_)) => {
+            CompositeEmbedderContainsHuggingFace::Both
+        }
+        (SubEmbedder::HuggingFace(_), _) => CompositeEmbedderContainsHuggingFace::Search,
+        (_, SubEmbedder::HuggingFace(_)) => CompositeEmbedderContainsHuggingFace::Indexing,
+        _ => CompositeEmbedderContainsHuggingFace::None,
+    }
+}
diff --git a/crates/milli/src/vector/error.rs b/crates/milli/src/vector/error.rs
index 650249bff..0993ded1d 100644
--- a/crates/milli/src/vector/error.rs
+++ b/crates/milli/src/vector/error.rs
@@ -6,6 +6,7 @@ use hf_hub::api::sync::ApiError;
 
 use super::parsed_vectors::ParsedVectorsDiff;
 use super::rest::ConfigurationSource;
+use super::MAX_COMPOSITE_DISTANCE;
 use crate::error::FaultSource;
 use crate::update::new::vector_document::VectorDocument;
 use crate::{FieldDistribution, PanicCatched};
@@ -335,6 +336,77 @@ impl NewEmbedderError {
     pub(crate) fn ollama_unsupported_url(url: String) -> NewEmbedderError {
         Self { kind: NewEmbedderErrorKind::OllamaUnsupportedUrl(url), fault: FaultSource::User }
     }
+
+    pub(crate) fn composite_dimensions_mismatch(
+        search_dimensions: usize,
+        index_dimensions: usize,
+    ) -> NewEmbedderError {
+        Self {
+            kind: NewEmbedderErrorKind::CompositeDimensionsMismatch {
+                search_dimensions,
+                index_dimensions,
+            },
+            fault: FaultSource::User,
+        }
+    }
+
+    pub(crate) fn composite_test_embedding_failed(
+        inner: EmbedError,
+        failing_embedder: &'static str,
+    ) -> NewEmbedderError {
+        Self {
+            kind: NewEmbedderErrorKind::CompositeTestEmbeddingFailed { inner, failing_embedder },
+            fault: FaultSource::Runtime,
+        }
+    }
+
+    pub(crate) fn composite_embedding_count_mismatch(
+        search_count: usize,
+        index_count: usize,
+    ) -> NewEmbedderError {
+        Self {
+            kind: NewEmbedderErrorKind::CompositeEmbeddingCountMismatch {
+                search_count,
+                index_count,
+            },
+            fault: FaultSource::Runtime,
+        }
+    }
+
+    pub(crate) fn composite_embedding_value_mismatch(
+        distance: f32,
+        hint: CompositeEmbedderContainsHuggingFace,
+    ) -> NewEmbedderError {
+        Self {
+            kind: NewEmbedderErrorKind::CompositeEmbeddingValueMismatch { distance, hint },
+            fault: FaultSource::User,
+        }
+    }
+}
+
+#[derive(Debug, Clone, Copy)]
+pub enum CompositeEmbedderContainsHuggingFace {
+    Both,
+    Search,
+    Indexing,
+    None,
+}
+
+impl std::fmt::Display for CompositeEmbedderContainsHuggingFace {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            CompositeEmbedderContainsHuggingFace::Both => f.write_str(
+                "\n  - Make sure the `model`, `revision` and `pooling` of both embedders match.",
+            ),
+            CompositeEmbedderContainsHuggingFace::Search => f.write_str(
+                "\n  - Consider trying a different `pooling` method for the search embedder.",
+            ),
+            CompositeEmbedderContainsHuggingFace::Indexing => f.write_str(
+                "\n  - Consider trying a different `pooling` method for the indexing embedder.",
+            ),
+            CompositeEmbedderContainsHuggingFace::None => Ok(()),
+        }
+    }
 }
 
 #[derive(Debug, thiserror::Error)]
@@ -419,6 +491,14 @@ pub enum NewEmbedderErrorKind {
     CouldNotParseTemplate(String),
     #[error("unsupported Ollama URL.\n  - For `ollama` sources, the URL must end with `/api/embed` or `/api/embeddings`\n  - Got `{0}`")]
     OllamaUnsupportedUrl(String),
+    #[error("error while generating test embeddings.\n  - the dimensions of embeddings produced at search time and at indexing time don't match.\n  - Search time dimensions: {search_dimensions}\n  - Indexing time dimensions: {index_dimensions}\n  - Note: Dimensions of embeddings produced by both embedders are required to match.")]
+    CompositeDimensionsMismatch { search_dimensions: usize, index_dimensions: usize },
+    #[error("error while generating test embeddings.\n  - could not generate test embedding with embedder at {failing_embedder} time.\n  - Embedding failed with {inner}")]
+    CompositeTestEmbeddingFailed { inner: EmbedError, failing_embedder: &'static str },
+    #[error("error while generating test embeddings.\n  - the number of generated embeddings differs.\n  - {search_count} embeddings for the search time embedder.\n  - {index_count} embeddings for the indexing time embedder.")]
+    CompositeEmbeddingCountMismatch { search_count: usize, index_count: usize },
+    #[error("error while generating test embeddings.\n  - the embeddings produced at search time and indexing time are not similar enough.\n  - angular distance {distance}\n  - Meilisearch requires a maximum distance of {MAX_COMPOSITE_DISTANCE}.\n  - Note: check that both embedders produce similar embeddings.{hint}")]
+    CompositeEmbeddingValueMismatch { distance: f32, hint: CompositeEmbedderContainsHuggingFace },
 }
 
 pub struct PossibleEmbeddingMistakes {
diff --git a/crates/milli/src/vector/mod.rs b/crates/milli/src/vector/mod.rs
index 74b52b1fe..47307295e 100644
--- a/crates/milli/src/vector/mod.rs
+++ b/crates/milli/src/vector/mod.rs
@@ -15,6 +15,7 @@ use self::error::{EmbedError, NewEmbedderError};
 use crate::prompt::{Prompt, PromptData};
 use crate::ThreadPoolNoAbort;
 
+pub mod composite;
 pub mod error;
 pub mod hf;
 pub mod json_template;
@@ -31,6 +32,7 @@ pub use self::error::Error;
 pub type Embedding = Vec<f32>;
 
 pub const REQUEST_PARALLELISM: usize = 40;
+pub const MAX_COMPOSITE_DISTANCE: f32 = 0.01;
 
 pub struct ArroyWrapper {
     quantized: bool,