Cohere launches Rerank 5 Pro and Fast for enterprise retrieval

Cohere describes Rerank 5 Pro as optimized for complex, reasoning-heavy queries and long or domain-specific documents.[2] The company says its largest gains for Pro are in reasoning-intensive retrieval and areas including finance, healthcare and scientific research.[2] Rerank 5 Fast is designed for latency- and cost-sensitive workloads, with optimization for low latency and high throughput.[2] Cohere says Fast outperforms the previous-generation Rerank 4 Pro, but the announcement does not provide a quantified comparison.[2][3]

Rerank 5’s Context Focus feature returns relevant passages within each document as scored character spans in the same reranking call.[2] Cohere says users can pass only selected passages to an LLM to reduce token costs, or highlight those passages to explain why a search matched.[2] The feature is disabled by default.[2] Cohere also says relevance scores are consistent across queries and domains, allowing users to set one threshold for deciding which results reach an LLM.[2]

Cohere says Rerank 5 supports more than 100 languages and has a 32,000-token context window.[2] It can also rerank semi-structured data such as JSON, tables and records formatted as YAML strings.[2] For customers moving from an earlier Rerank model, Cohere says no re-indexing or re-embedding is needed because Rerank reads raw text at query time; upgrading involves changing the model parameter to either Rerank 5 variant.[2]

Cohere says Rerank 5 is available through its Rerank API, with both Pro and Fast also available through Microsoft Foundry and Amazon SageMaker.[2] AI Matter reported that Cohere provided no test numbers, no figure for how much text Context Focus cuts, and no quantified improvement over Rerank 4.[3]