AI Potluck
Product / UX / Agent tools & protocols

Cohere Rerank API

Cohere

Reranking API that takes a query and a list of documents and returns them reordered by relevance. It is built for RAG pipelines: an agent retrieves candidates from a vector database, then calls Rerank to pick the most relevant before passing them to a model. It is offered standalone rather than bundled into a vector database, and applied as a single API call.

The declared pypi artifact is Cohere's whole client SDK rather than a reranker package, so it cannot measure this product; adoption rests on multi-cloud distribution instead. Every published performance figure is vendor-run, which is why capability stops at 4. Verified 2026-08-13 via the Cohere rerank overview, the 3.5 changelog and the 3.5 launch post.

Openness

1 high confidence
1.0
license
proprietary(hosted-API
source
closed
weights
not-released
managed
Cohere+Azure/AWS/Oracle

Proprietary closed reranking models served via API and cloud marketplaces; no open weights or source. Marketed as 'open' in search-relevance contexts but is not OSS.

  • https://docs.cohere.com/docs/rerank-overview recorded 2026-08-13

    Model lineup - "Rerank 4.0 (both 'fast' and 'pro'): A single multilingual model (rerank-v4.0-pro and rerank-v4.0-fast). Rerank 3.5: A single multilingual model (rerank-v3.5). Rerank 3.0: Separate English-only and multilingual models." Calls are billed in search_units. No weights download, hub model card or self-hosted build for any of them.

  • https://docs.cohere.com/changelog/rerank-v3.5 recorded 2026-08-13

    Release notes for Rerank 3.5 as a hosted model reached through the API, with a context length of 4096. No weights release accompanies it.

Adoption

3 medium confidence
3.0

The de facto default managed reranker for RAG, distributed through Azure AI Foundry, AWS SageMaker JumpStart and OpenSearch, and Oracle GenAI as well as Cohere's own API. No call-volume or user count is published anywhere, so the level rests on that breadth of multi-cloud distribution rather than on a measured figure.

Capability

4 medium confidence
4.0

Cohere reports state-of-the-art results on BEIR and on multilingual retrieval, and the models are widely treated as the reference managed reranker. Both headline figures are vendor-run against vendor baselines, though: the BEIR claim appears in the Rerank 3.5 release notes, and the +26.4% cross-lingual gain is measured against Cohere's own previous Rerank 3 on an in-house 18-language suite scored by nDCG@10. No concrete nDCG numbers are disclosed on the primary pages, and no standard public reranker leaderboard applies, so the score rests on a comparison of features and stops at 4 rather than 5.

  • https://cohere.com/blog/rerank-3pt5 recorded 2026-08-13

    "Cohere's multilingual evaluation suite consists of external datasets covering 18 different languages in a variety of monolingual and cross-lingual settings. Multilingual performance is measured by nDCG@10"; "When compared to our previous Rerank 3 model, Rerank 3.5 delivers a +26.4% improvement on cross-lingual search". NOT on this page: BEIR, or any comparison against embedding-only retrieval.

  • https://docs.cohere.com/changelog/rerank-v3.5 recorded 2026-08-13

    "Rerank 3.5 has a context length of 4096, SOTA performance on Multilingual Retrieval tasks and Reasoning Capabilities. In addition, Rerank 3.5 has SOTA performance on BEIR and domains such as Finance, E-commerce, Hospitality, Project Management, and Email/Messaging Retrieval tasks." No nDCG figure is given here.

Verified 2026-08-13