A passage can answer a question and omit the exception that changes its meaning. Keep the original document within reach when testing shorter context.
What Cohere released on October 21
Cohere’s release note introduces rerank-v5.0-pro and rerank-v5.0-fast. It lists access through the Rerank API, Microsoft Foundry, Amazon SageMaker, and Model Vault. Private deployments require a sales discussion.
Context Focus can return relevant passages as character ranges within each supplied document. It is disabled unless requested. These spans let an application select evidence before building the next prompt.
An upgrade and a context change need separate checks
Cohere says reranking reads documents at query time, so changing the model does not require rebuilding embeddings. Sending only selected passages creates another change: the downstream model receives less evidence.
I would evaluate those changes separately. First, compare ranking with the full document context preserved. Then enable passage selection and measure how the final answers change.
Cohere’s quality and score-calibration statements are vendor claims. A relevance score estimates a retrieval relationship; it cannot establish that an answer is factually correct.
Build a test around missing exceptions
Use questions with known answers and source locations. Include a policy with an exception in another paragraph, a table with a footnote, and two documents that disagree.
- Keep the query, document versions, and expected answer fixed.
- Save the returned document order and selected character ranges.
- Check whether a required exception survives passage selection.
- Record unsupported answers and refusals alongside successful answers.
- Compare latency and total retrieval-plus-generation cost.
If an answer needs a nearby qualifier, try a wider passage or retain the complete document for that question class. Review the result again on held-out examples.
Count the whole request
A shorter prompt can reduce generation input charges. The rerank call adds latency, and unsuccessful answers still count in the comparison.
Divide total evaluation spending by accepted answers. Record human correction time separately. This lets you reject a cheaper context policy when it creates more review work.
Our FinRank article explains why a correct answer can cite the wrong evidence. The AI Search guide covers a separate retrieval service and its indexing costs.
Read the sources and choose your next check
Which exception would you include in the first passage-selection test?
Checked October 21, 2026. This article is based on source review and analysis. Suggested tests have not been performed by MustHave.ai.
