Struct MultiVectorRetriever 

pub struct MultiVectorRetriever<E: MultiVectorEmbedder> { /* private fields */ }

Multi-vector retriever using WARP index for ColBERT-style late interaction.

This retriever uses token-level embeddings and MaxSim scoring for fine-grained semantic matching. Unlike single-vector dense retrieval, multi-vector approaches represent documents and queries as multiple token embeddings.

Example

use trueno_rag::multivector::{
    WarpIndexConfig, WarpSearchConfig,
    MockMultiVectorEmbedder, MultiVectorRetriever,
};

let config = WarpIndexConfig::new(2, 256, 128);
let embedder = MockMultiVectorEmbedder::new(128, 512);
let mut retriever = MultiVectorRetriever::new(config, embedder);

// Train on sample documents
retriever.train(&sample_chunks)?;

// Index documents
for chunk in chunks {
    retriever.index(chunk)?;
}
retriever.build()?;

// Search
let results = retriever.retrieve("What is machine learning?", 10)?;

Implementations

impl<E: MultiVectorEmbedder> MultiVectorRetriever<E>

pub fn new(config: WarpIndexConfig, embedder: E) -> Self

Create a new multi-vector retriever with the given configuration and embedder.

Arguments

• config - WARP index configuration (nbits, num_centroids, token_dim)
• embedder - Multi-vector embedder for generating token embeddings

pub fn with_search_config(self, config: WarpSearchConfig) -> Self

Set the search configuration.

pub fn train(&mut self, sample_chunks: &[Chunk]) -> Result<()>

Train the WARP index on sample chunks.

This builds the residual quantization codec by learning centroids from the provided sample embeddings. Should be called before indexing.

Arguments

• sample_chunks - Representative chunks for training the codec

pub fn index(&mut self, chunk: Chunk) -> Result<()>

Index a single chunk.

The chunk is embedded and compressed using the trained codec. Call train() before indexing.

pub fn index_batch(&mut self, chunks: Vec<Chunk>) -> Result<()>

Index multiple chunks.

pub fn build(&mut self) -> Result<()>

Build the index for efficient search.

This compacts the index by organizing embeddings by centroid (IVF structure). Call after all chunks have been indexed.

pub fn retrieve(&self, query: &str, k: usize) -> Result<Vec<RetrievalResult>>

Retrieve relevant chunks for a query using multi-vector MaxSim scoring.

Arguments

• query - Query text
• k - Number of results to return

pub fn len(&self) -> usize

Get the number of indexed chunks.

pub fn is_empty(&self) -> bool

Check if the retriever is empty.

pub fn warp_index(&self) -> &WarpIndex

Get the underlying WARP index.

pub fn embedder(&self) -> &E

Get the embedder.

pub fn memory_usage(&self) -> usize

Get memory usage statistics.