Cohere released Rerank 4, a new set of reranker models for enterprise AI search, with claims of stronger retrieval relevance than MongoDB’s Voyage models and ElasticSearch’s Jina rerankers, lower latency, flexible deployment, customization, and multilingual performance. The company positions Rerank 4 for business-critical search in finance, healthcare, manufacturing, e-commerce, programming, and customer service, where users need documents, passages, product listings, or internal knowledge ranked by relevance.

Rerankers improve enterprise search after an initial retrieval step, such as BM25 or bi-encoder embeddings, produces a broad list of candidate results. Rerank 4 uses a cross-encoder architecture (jointly reading query and result) to compare queries with candidates together, capture semantic relationships, and reorder results without evaluating an entire corpus.

Rerank 4 is part of North, Cohere’s agentic AI platform, alongside Embed, Rerank, the Command model series, and customizable AI agents. Cohere says the model works with hybrid, vector, and keyword-based systems with minimal code changes, and can improve retrieval-augmented generation (RAG) pipelines by filtering irrelevant content before it reaches a generative model, reducing token usage and costly retries in multi-step agent workflows.

Rerank 4 has a 32K context window, the largest in Cohere’s Rerank series and a four-fold increase over the previous generation. Cohere says the longer context lets the model handle longer documents, evaluate multiple passages at once, and capture relationships across sections that shorter windows would miss.

Cohere offers Rerank 4 in two versions: Fast and Pro. Fast is a smaller model for workloads needing both speed and accuracy, including e-commerce product recommendations, semantic catalogue search, engineering documentation lookup, and customer-service ticket triage. Pro is optimized for tasks requiring deeper reasoning, analysis, and higher precision, including finance risk models using market reports, regulatory filings, and transaction histories; healthcare review of patient records and clinical trial reports; and manufacturing diagnosis using technical manuals and production logs.

Cohere says Rerank 4 outperforms current competitive alternatives across enterprise domains including finance, healthcare, manufacturing, and workplace productivity. Its evaluations use retrieval performance measured by nDCG@10, including public BEIR benchmarks, internal long-context benchmarks, parsed PDF datasets across financial, healthcare, biomedical, and healthcare domains spanning 6 languages, and semi-structured retrieval over mixtures of metadata and free-form text.

Rerank 4 also supports deployment in over 100 world languages, including state-of-the-art retrieval in 10 major business languages: Arabic, Chinese, French, German, Hindi, Japanese, Korean, Portuguese, Russian, and Spanish. Cohere says its multilingual evaluation suite covers 18 different languages in monolingual and cross-lingual settings, with multilingual performance measured by nDCG@10.

Rerank 4 is Cohere’s first reranking model with self-learning capability. In partnership with Cohere, users can customize the model for specific use cases without further annotated data by stating preferences for content types, prescribed language or terminology, or document corpora. Cohere says Self Learning helped Rerank 4 Fast improve iteratively, become competitive with larger market alternatives, and in many tests converge with or surpass out-of-the-box Pro performance.

Cohere tested Self Learning on ViDoRe V3, a public benchmark containing visually rich, multimodal enterprise documents such as reports, invoices, forms, and technical documentation. The company also used healthcare-focused datasets that mimic clinician retrieval of patient-specific information and reported consistent, substantial gains for Rerank 4 Fast.

Rerank 4 is available on Cohere’s Platform, Amazon SageMaker AI for Fast and Pro, and Microsoft Foundry, with more platform support planned. Cohere says Rerank 4 can also be deployed in any Virtual Private Cloud or on-premise environment for large enterprise deployments.

Cohere released Rerank 4, a new set of reranker models for enterprise AI search, with claims of stronger retrieval relevance than MongoDB’s Voyage models and ElasticSearch’s Jina rerankers, lower latency, flexible deployment, customization, and multilingual performance. The company positions Rerank 4 for business-critical search in finance, healthcare, manufacturing, e-commerce, programming, and customer service, where users need documents, passages, product listings, or internal knowledge ranked by relevance.

Rerankers improve enterprise search after an initial retrieval step, such as BM25 or bi-encoder embeddings, produces a broad list of candidate results. Rerank 4 uses a cross-encoder architecture (jointly reading query and result) to compare queries with candidates together, capture semantic relationships, and reorder results without evaluating an entire corpus.

Rerank 4 is part of North, Cohere’s agentic AI platform, alongside Embed, Rerank, the Command model series, and customizable AI agents. Cohere says the model works with hybrid, vector, and keyword-based systems with minimal code changes, and can improve retrieval-augmented generation (RAG) pipelines by filtering irrelevant content before it reaches a generative model, reducing token usage and costly retries in multi-step agent workflows.

Rerank 4 has a 32K context window, the largest in Cohere’s Rerank series and a four-fold increase over the previous generation. Cohere says the longer context lets the model handle longer documents, evaluate multiple passages at once, and capture relationships across sections that shorter windows would miss.

Cohere offers Rerank 4 in two versions: Fast and Pro. Fast is a smaller model for workloads needing both speed and accuracy, including e-commerce product recommendations, semantic catalogue search, engineering documentation lookup, and customer-service ticket triage. Pro is optimized for tasks requiring deeper reasoning, analysis, and higher precision, including finance risk models using market reports, regulatory filings, and transaction histories; healthcare review of patient records and clinical trial reports; and manufacturing diagnosis using technical manuals and production logs.

Cohere says Rerank 4 outperforms current competitive alternatives across enterprise domains including finance, healthcare, manufacturing, and workplace productivity. Its evaluations use retrieval performance measured by nDCG@10, including public BEIR benchmarks, internal long-context benchmarks, parsed PDF datasets across financial, healthcare, biomedical, and healthcare domains spanning 6 languages, and semi-structured retrieval over mixtures of metadata and free-form text.

Rerank 4 also supports deployment in over 100 world languages, including state-of-the-art retrieval in 10 major business languages: Arabic, Chinese, French, German, Hindi, Japanese, Korean, Portuguese, Russian, and Spanish. Cohere says its multilingual evaluation suite covers 18 different languages in monolingual and cross-lingual settings, with multilingual performance measured by nDCG@10.

Rerank 4 is Cohere’s first reranking model with self-learning capability. In partnership with Cohere, users can customize the model for specific use cases without further annotated data by stating preferences for content types, prescribed language or terminology, or document corpora. Cohere says Self Learning helped Rerank 4 Fast improve iteratively, become competitive with larger market alternatives, and in many tests converge with or surpass out-of-the-box Pro performance.

Cohere tested Self Learning on ViDoRe V3, a public benchmark containing visually rich, multimodal enterprise documents such as reports, invoices, forms, and technical documentation. The company also used healthcare-focused datasets that mimic clinician retrieval of patient-specific information and reported consistent, substantial gains for Rerank 4 Fast.

Rerank 4 is available on Cohere’s Platform, Amazon SageMaker AI for Fast and Pro, and Microsoft Foundry, with more platform support planned. Cohere says Rerank 4 can also be deployed in any Virtual Private Cloud or on-premise environment for large enterprise deployments.