Reranking

Reranking applies a more precise scoring method to a small set of retrieved candidates, then reorders them before context is assembled.¹

Tiny example

A fast retriever finds 20 passages. A cross-encoder scores each query–passage pair and sends the best five to the language model.

20 fast candidates → precise reranker → 5 context passages

Important limit

A reranker can reorder only what the first retriever found. It cannot recover a relevant passage missing from the candidate set.

FDE note

Measure end-to-end answer quality and latency. Better ranking metrics are useful only if the added model call improves the customer outcome within the response-time budget.

Footnotes

1. Cohere’s Rerank overview describes the query-plus-documents interface and returned relevance ordering. Other providers may implement the same pattern differently. ↩