Vector search returns the nearest chunks, not the best ones. A reranker reads the candidates with the question and reorders them, which is where most of the answer quality in retrieval-augmented generation comes from. This release adds the stage and the chain around it.
Rerankers
- One interface A reranker contract with a result limit, so the model behind it can change without touching the chain.
- Cohere A reranker against Cohere's rerank API with retries.
- Local A reranker against a local Infinity service, with a timeout, for estates that keep retrieval on premises.
The chain
- Retrieve, rerank, answer A question-answer chain wires the top-N retriever, the reranker and a query transformer that rewrites the question before retrieval.
- Parallel embedding Embedding runs in parallel with retries, and the context retriever takes a configurable context window.
- Cohere chat A Cohere chat model joins the provider list.
Vertex AI reranking and multi-query retrieval followed two years later on the same chain.