最新模型
Cohere Releases Rerank 5, Adding Passage Extraction and Adjustable Thresholds to Reduce RAG Context
On October 21, Cohere released Rerank 5 Pro and Fast, adding Context Focus, which extracts relevant passages from documents during reranking. Developers can adjust the passage score threshold to control how much text is sent to a language model, but should still verify that the trimmed evidence remains complete.
Cohere released Rerank 5 Pro and Fast on October 21, extending post-retrieval processing from document ranking to passage selection within documents. The new Context Focus feature identifies relevant text in the same reranking call, so RAG systems and search agents can send only selected content to downstream models. Both models are available through channels including the Cohere API. Official announcement
This change directly affects how retrieval pipelines assemble context. Previously, systems often used an entire chunk after retrieving candidate documents; they can now use passage positions in the original text to build a more concise evidence set. A potential engineering benefit is reducing downstream input volume, but the savings depend on document length, query type, and the retention threshold. They should not be assumed to translate into a proportional reduction in overall costs.
According to the official documentation, Context Focus is disabled by default, and requests must include relevant_spans. Returned passages include their start and end character positions in the original text, along with a score. Raising threshold removes more text, while max_spans limits the number of passages retained per document. If no passage meets the threshold, the result omits the passage field, so integrations need to handle that case. Cohere also cautions that broad queries such as summarization usually require the full document, while segmentation can be less reliable for tables, forms, and code. Technical documentation
Cohere staff confirmed in a release discussion in the LangChain community that the passage threshold can be adjusted per request, and said the models are available through LangChain. This gives applications the option to vary the degree of trimming by query—for example, reducing input for precise question answering while retaining more evidence for questions that require synthesis across passages. Release discussion
On performance, Cohere reports that Pro ranked first on 39 of 46 diverse datasets, using the company’s RCP-nDCG@10 evaluation method. Those results should not be directly mixed with scores from other leaderboards. Chinese-language technical teams should next compare answer accuracy, citation completeness, end-to-end latency, and total usage on their own corpora, paying particular attention to whether trimming drops negations, exceptions, or dependencies across passages. Evaluation details