Skip to content
effect-uai

Reranking

The top-K vectors by cosine are not always the top-K results.

Embedding cosine compresses the relationship between a query and a document into one number per side. That’s good enough to filter a million documents down to fifty candidates. It’s often not good enough to pick the three the user actually wants. Rerankers re-score those fifty with a cross-encoder that sees the query and document together, trading throughput for accuracy where it matters.

The shape is small: one query, a list of documents, scores back. Same HTTP plumbing as embeddings, different model class.

Coming soon

@effect-uai/core will ship a Reranker service tag and provider implementations, planned in this rough order:

  • Jina — jina-reranker-v3 and jina-reranker-m0 (multimodal, multilingual).
  • Cohere — rerank-v3.5 and successors.
  • Voyage — rerank-2 and rerank-2-lite.
  • Mixedbread — mxbai-rerank-large-v2.

The four providers’ wire shapes already converge on POST /rerank { model, query, documents } → { results: [...] }, so one common service tag covers all of them with provider-typed tags extending it for per-provider knobs.

Show interest

Want this sooner, or have a provider you’d like to see prioritised? Open or +1 the reranking tracking issue.