Module score 

Host-side similarity scoring for sparse and multi-vector embeddings.

uni-xervo emits sparse and multi-vector embeddings but deliberately does not own an index — native multi-vector / inverted-index storage is a downstream concern. These pure, dependency-free helpers close that gap for the common case of reranking a small top-k candidate set in the host process:

• max_sim / colbert_rerank score per-token (ColBERT) vectors via late interaction (sum of per-query-token maxima).
• sparse_dot scores two learned-sparse vectors.

They make these embeddings useful immediately, without waiting on a native index. For large corpora, prefer a real index over scoring every candidate here.

Functions

colbert_rerank

Rank docs against query by max_sim, returning one score per document.

max_sim

Compute the MaxSim late-interaction score between query and document tokens.

sparse_dot

Compute the dot product of two learned-sparse vectors.