Refine your enterprise search and retrieval with precise result ranking

A reranker is the second stage of a search pipeline. After first-stage search retrieves a shortlist, the reranker reads the query alongside each result and reorders them so the most relevant come first.

Yes, if accuracy matters. Rerank 5 reads the query and each result together, so it catches details that vector search can miss, such as numbers, dates and negation.

It is slower and costs more. Ranking 100 results with an LLM takes several sequential calls, and every irrelevant token you send costs money and can lower answer quality. Rerank does the job in one call.

Pro is the most accurate, for complex queries and specialist documents. Fast is built for speed and volume, such as interactive search and agents, and still outperforms many leading market alternatives.

Context Focus is an optional Rerank 5 feature that returns the most relevant passages in each document, in the same call. Sending only those passages to your LLM cuts tokens with minimal loss in answer quality.

Rerank 5's scores mean roughly the same thing across queries, so one threshold works everywhere. Use it to drop weak matches, let an assistant say "I don't know" or trigger another search.

RCP-nDCG@10 is Cohere's method for scoring the top 10 search results, and the metric used to evaluate Rerank 5 and Embed 5. Standard nDCG credits only the documents in a benchmark's answer key. RCP-nDCG instead grades every result against the same relevance criteria, and it has been validated against blind human judgments.