Finding the right contract does not necessarily give an AI assistant the right evidence. A termination clause may depend on a notice period or an exception elsewhere in the section. Cohere’s Rerank 5, announced on 21 October, adds passage selection to the task of ranking retrieved documents.[1][2]
Two stages of selecting evidence
A reranker reviews candidates found by an initial keyword or vector search and reorders them against the question. Rerank 5 comes in Pro and Fast variants. Its optional Context Focus feature returns scored character spans within those documents, allowing an application to select passages or highlight why a result matched. It is off by default.[2]
Cohere reports that Pro ranked first on 39 of 46 public composite datasets. On BRIGHT, its score was 82.2, ahead of Voyage AI Rerank 3 by 5.8 points under Cohere’s RCP-nDCG@10 methodology. That measures the ordering of relevant results near the top of a list; it is not the percentage of final answers that are factually correct.[1]
The answer-level experiment
Cohere’s search-agent comparison
| Measure | Value | Context |
|---|---|---|
| Keyword search alone | 46% answer accuracy | |
| With Rerank 5 Pro | 49% answer accuracy | |
| With Rerank 5 Pro and Context Focus at 0.4 | 58% answer accuracy |
The token measure includes input, output, reasoning and cached tokens across every model call. It is therefore broader than the length of the retrieved passages. In a separate single-turn test across 30 datasets, removing half the retrieved context reduced the average model-judged answer-quality score by two percentage points. These experiments measure different workflows and should not be combined into one universal saving.[1]
What must survive the filter
The risk is selecting the right document but dropping the sentence that controls its meaning. Consider an illustrative rule: contractors can access a system except while under investigation. Keeping the permission and losing the exception gives an answering model a misleading fragment even though retrieval found the correct policy.
That makes passage selection an end-to-end question: does the final answer preserve dates, thresholds and exceptions, and can the reader inspect the surrounding source? A smaller context is useful when it removes distraction while keeping the evidence that changes the answer.
The product documentation says upgrades do not require re-embedding the corpus because reranking reads raw text at query time. Teams can therefore compare the new ranking and passage-selection stages against their existing retrieval flow without rebuilding the underlying index.[2]
Sources & further reading
- Introducing Rerank 5: precise, efficient enterprise document retrieval Cohere · 21 Oct 2026 · Accessed 26 Oct 2026
- Rerank v5 product release documentation Cohere Documentation · 21 Oct 2026 · Accessed 26 Oct 2026