> ## Content Index > Fetch the complete content index at: https://www.tao.media/llms.txt > Use this file to discover other available public pages before exploring further. # Cohere Launches Rerank 5 With Passage-Level Retrieval for AI Agents - URL: https://www.tao.media/cohere-launches-rerank-5-with-passage-level-retrieval-for-ai-agents/ - Published: 2026-10-23T17:57:59.000Z - Updated: 2026-10-23T17:57:59.000Z - Description: Context Focus selects relevant passages within documents to reduce the context AI agents need to process. - Author: Tristan Hillerich - Tags: Cohere, AI, News Cohere has [launched Rerank 5](https://x.com/cohere/status/2112923503100125326?ref=tao.media), a family of enterprise search models that ranks retrieved documents and identifies the passages an AI system needs to answer a question. Available in Pro and Fast tiers, the release adds Context Focus, an optional feature for sending selected evidence to a language model instead of entire documents. In Cohere's search-agent evaluations, a retrieval stack using Rerank 5 Pro and Context Focus consumed 35% fewer tokens than keyword search alone while improving answer accuracy. The reduction reached 41% across six deep-search benchmarks, where accuracy rose from 39% to 52%. Both models are generally available through the Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker, and Cohere's North enterprise AI platform. Pro prioritizes ranking quality, while Fast offers lower pricing and higher throughput for applications that search repeatedly. ## How Context Focus Selects Evidence A reranker operates after an initial search. A keyword, vector, or hybrid system retrieves candidate documents, and the reranker evaluates how useful each candidate is for the query. In retrieval-augmented generation, or RAG, those results become evidence for a language model's answer. Document-level ranking can still leave an application sending substantial irrelevant text downstream. A financial report might contain the answer in one sentence, surrounded by pages about unrelated business lines. Context Focus adds scored text spans to the document-level relevance scores returned by a single API call, allowing the application to select passages rather than forward the full document. Developers can adjust a threshold to control how much text survives. In [Cohere's evaluations across 30 question-answering datasets](https://cohere.com/blog/rerank-5?ref=tao.media), removing half the retrieved context retained 96% of answer quality, while removing three-quarters retained 87%. The experiment used Qwen3.5-9B to answer questions and Qwen3.5-35B to grade the responses. The separate search-agent evaluation measured the effect across a multi-step workflow. Cohere tested 17 datasets, including six deep-search benchmarks, using glm-5-3-flash with search as its only tool. The agent received five results per search, with Rerank 5 Pro reordering the top 100 candidates from BM25 keyword search. ![Context Focus: search-agent accuracy vs. tokens per task](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/10/HVKaBrKWIAALfQZ.png) Context Focus: search-agent accuracy vs. tokens per task | [Cohere on X](https://x.com/cohere/status/2112923515393290528?ref=tao.media) Adding Pro without Context Focus increased accuracy from 46% to 49%. Enabling Context Focus at its default 0.1 threshold raised accuracy to 57% and reduced tokens per task by 19% relative to reranking without passage filtering. At a 0.4 threshold, accuracy reached 58%, and the complete stack used 35% fewer tokens than BM25 alone. Those token measurements cover the entire agent task: input, output, and reasoning tokens across all language-model calls, including cached tokens. They are not simply a measure of the retrieved text removed. Cohere used 50 questions per dataset and reported improvements in both accuracy and token use on every deep-search benchmark. ## Relevance Scores Become Usable Cutoffs Rerank 5 also changes how applications can interpret relevance scores. Cohere has calibrated the scores to approximate the probability that a document is useful. A score of 0.7 corresponds to roughly a 70% likelihood of usefulness. That gives developers an alternative to always forwarding a fixed number of top-ranked documents. A narrow query can pass a small set of high-scoring results, while a broader question can retain more. When nothing clears the threshold, an agent can search again rather than automatically use the highest-ranked result. Cohere evaluated calibration against two sets of human judgments: a blind panel covering 2,262 documents and NIST TREC Deep Learning assessments covering 20,646 documents. Pro's expected calibration error was 0.058 on the panel and 0.057 on NIST; Fast scored 0.058 on both. This metric measures the average gap between predicted relevance and the observed share of useful documents, with lower values indicating closer agreement. The company reported errors of 0.234 and 0.307 for Voyage AI Rerank 3 on the same two sets. Pro and Fast also ranked first and second in average precision in those evaluations. Their usefulness cutoffs were close, though not identical: 0.50 for Pro and 0.45 for Fast represented documents judged more likely useful than not. ## Pro Leads Cohere's Tests, While Fast Targets Volume Cohere reports that Rerank 5 Pro outperformed every competing model it tested across ViDoRe V3, BEIR, BRIGHT, and nanoMTEB. Pro ranked first on 39 of the 46 constituent datasets and at least second on 43. The results use RCP-nDCG@10, a Cohere-developed ranking evaluation methodology that addresses incomplete relevance labels. The scores assess how well useful results are ordered near the top of a retrieval list. ![Rerank 5 headline benchmarks: Pro and Fast vs. competitors on ViDoRe V3, BEIR, BRIGHT, and nanoMTEB](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/10/HVKZ2dAXkAAMUhJ.jpg) Rerank 5 headline benchmarks: Pro and Fast vs. competitors on ViDoRe V3, BEIR, BRIGHT, and nanoMTEB | [Cohere on X](https://x.com/cohere/status/2112923508108132427?ref=tao.media) On BRIGHT, which tests retrieval requiring reasoning beyond surface-level similarity, Pro scored 82.2, beating Voyage AI Rerank 3 by 5.8 points and Rerank 4 Pro by 12.4 points. Its leads over the strongest external competitor were 1.6 points on ViDoRe V3, 1.8 on BEIR, and 3.4 on nanoMTEB. The largest reported domain-level advantage was in finance and economics, where Pro led Voyage AI Rerank 3 by 9.6 points across the evaluated tasks. In Cohere's internal multilingual suite, Pro ranked first in 14 of 19 non-English languages. Across 47 datasets of text parsed from PDFs, it averaged 89.9, compared with 88.4 for Voyage AI Rerank 3 and 88.3 for Rerank 4 Pro. Fast also outperformed Voyage AI Rerank 3 Lite, Qwen3-Reranker-0.6B, Qwen3-Reranker-8B, and the previous Rerank 4 Pro across all four headline benchmark averages in Cohere's tests. ![Rerank 5 Fast vs. Pro throughput and latency](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/10/HVKZ9O0XAAAcQxq.jpg) Rerank 5 Fast vs. Pro throughput and latency | [Cohere on X](https://x.com/cohere/status/2112923512092676477?ref=tao.media) For 1,000-token documents, Fast processed about 187 documents per second on a single NVIDIA A100 GPU, delivering 2.9 times Pro's throughput. Cohere also reports nearly threefold faster response times. Its throughput and median-latency tests used 96 documents per request, concurrency of eight, and 1,000 requests per document length.