Module multivector 

Multi-vector (late-interaction / ColBERT) embedding types and the MultiVectorEmbeddingModel trait.

Where EmbeddingModel pools a sequence into one dense vector, a multi-vector model emits one vector per token. Scoring is then done with MaxSim (sum over query tokens of the max similarity to any document token) — the late-interaction approach popularized by ColBERT and the strongest known method for layout-rich document retrieval (ColPali / ColQwen2).

This trait is a separate, narrow capability so a provider opts into it only when it can skip pooling and surface per-token vectors. The native multi-vector index is deliberately out of scope — this is the producer side, immediately usable via host-side MaxSim (see max_sim).

Structs

MultiVectorEmbedResult

Result of a multi-vector embedding call: per-token vectors for each input.

Traits

MultiVectorEmbeddingModel

A model that produces per-token (multi-vector / ColBERT) embeddings from text.