Cohere released Rerank 5 on October 21, tackling a practical problem: deciding which information an AI system should read before answering. Pro targets complex searches; Fast targets low-latency applications and agents that search repeatedly. Both are available through Cohere’s API, Microsoft Foundry, and Amazon SageMaker.
Context Focus goes inside the document
The standout addition is Context Focus. Rather than only ranking documents, it can identify relevant passages inside them. Developers can send those passages to a language model instead of forwarding entire documents.
The optional feature is off by default and available through Cohere’s API. It returns scored character spans that point back into the original document, letting an application select context or highlight why a result matched.
A retrieval upgrade, not a new index
Rerank operates on text at query time, on top of existing search results. Existing Rerank users can switch model IDs without rebuilding or re-embedding their document index. That makes it a targeted component to evaluate within an existing retrieval pipeline, rather than a requirement to replace the whole search system.
Why it matters for builders
- Less irrelevant context: passage selection can reduce downstream token usage, although savings and answer quality depend on the workload.
- A choice of tradeoffs: Pro emphasizes complex relevance judgments, while Fast suits repeated, latency-sensitive searches.
- An important limitation: specific questions suit passage selection; broad summaries still need full documents. Tables, forms, and code also split less cleanly than prose.
A bounded SunMarc evaluation
For a potential SunMarc app-support assistant, this suggests a useful test: retrieve the right section of the QR Remix user guide instead of sending the whole guide.
Compare complete, correct answers with the amount of context used, checking that passage selection does not leave out a necessary step or caveat. This is a proposed evaluation, not an announced assistant feature or a claim of tested savings. The measure of success is complete answers with less context—not just better retrieval scores.