Struct MultiVectorSearcher 

pub struct MultiVectorSearcher { /* private fields */ }

A loaded multi-vector index, ready for MaxSim search.

Implementations

impl MultiVectorSearcher

pub fn open(index_path: &Path) -> Result<Self>

Open a multi-vector index from disk.

pub fn num_docs(&self) -> usize

Number of documents in the index.

pub fn num_tokens(&self) -> usize

Total number of token vectors in the index.

pub fn search( &self, query_tokens: &Array2<f32>, top_k: usize, ) -> Result<Vec<MultiVectorResult>>

Approximate MaxSim search.

For each query token, runs HNSW ANN search, then aggregates per-document using the MaxSim formula.

query_tokens has shape [num_query_tokens, dim].

pub fn search_with_params( &self, query_tokens: &Array2<f32>, top_k: usize, per_token_k: usize, ) -> Result<Vec<MultiVectorResult>>

Approximate MaxSim search with configurable per-token k.

pub fn search_exact( &self, query_tokens: &Array2<f32>, top_k: usize, first_stage_k: usize, ) -> Result<Vec<MultiVectorResult>>

Two-stage exact MaxSim search.

Stage 1: approximate HNSW search to find candidate doc_ids. Stage 2: exact MaxSim reranking using mmap’d embeddings.