Module sparse 

Learned-sparse text embedding types and the SparseEmbeddingModel trait.

Sparse embeddings represent text as a small set of weighted vocabulary terms rather than a single dense vector. They power hybrid first-stage retrieval: a learned-sparse model (SPLADE family, the BGE-M3 sparse head) keeps the inverted-index compatibility of BM25 while learning term weights — and term expansion, for SPLADE — that beat BM25 in and out of domain.

This trait sits alongside EmbeddingModel: it is a separate, narrow capability so a provider opts into it only when it can produce term-weight vectors. The downstream inverted index and impact scoring are deliberately out of scope — this is the producer side only.

Structs

SparseEmbedResult

Result of a sparse-embedding call, carrying one SparseVector per input.

Traits

SparseEmbeddingModel

A model that produces learned-sparse term-weight vectors from text.

Type Aliases

SparseVector

A single learned-sparse vector as (term_id, weight) pairs.