Key Contributors: Clifton Poth (Member of Technical Staff), Fabian David Schmidt (Member of Technical Staff), Martin Hentschel (Member of Technical Staff), Daniel Simig (Manager of Technical Staff), Nils Reimers (VP of Embeddings & Search), Elliott Choi (Director of Product)

Rerank 4 is the most advanced set of reranker models available today, purpose-built to meet the realities and challenges of enterprise AI search. It delivers best-in-class retrieval, outperforming the likes of MongoDB’s Voyage models and ElasticSearch’s Jina rerankers in overall search relevance, as well as improved latency, flexible deployment options, deep customizability and robust multilingual performance. Designed for business-critical applications across key industries and domains, Rerank 4 sets a new standard for accuracy and adaptability in enterprise search.

Why reranking matters

Rerankers significantly enhance the accuracy of enterprise AI search by refining initial retrieval results. After a fast but broad candidate-generation step using methods like BM25 or bi-encoder embeddings, the system is left with documents, passages, or product listings that approximate relevance but often miss the nuance of the user’s intent. Rerank 4 addresses this gap using a cross-encoder architecture that processes queries and candidates jointly, capturing subtle semantic relationships and reordering results to surface the most relevant items. This approach balances speed and accuracy, delivering high-quality results without evaluating the entire corpus.

Rerank 4 is a key component of North, Cohere’s agentic AI platform, which combines intelligent search (Embed, Rerank), large language models (our Command series), and customizable AI agents to automate tasks and accelerate decision-making. It integrates seamlessly into existing AI search solutions, including hybrid, vector, and keyword-based systems, with minimal code changes. Beyond improving retrieval-augmented generation (RAG) pipelines, Rerank 4 is critical in agentic AI systems, providing distilled, high-quality information for stronger reasoning and context-aware agent performance. By filtering out irrelevant content before it reaches the generative model, Rerank 4 reduces token usage and minimizes the number of costly retries an agent might otherwise need to “get it right,” sparing the user both pennies and latency. This is especially impactful for agentic AI, where complex, multi-step interactions can quickly drive up model calls and saturate context windows.

Rerank 4 Fast & Pro

Rerank 4 boasts a 32K context window - the largest of our Rerank series to date and a four-fold increase over the previous generation. This enables the model to handle longer documents, evaluate multiple passages simultaneously, and capture relationships across sections that shorter windows would miss. This expanded capacity therefore improves ranking accuracy for realistic document types and increases confidence in the relevance of retrieved results.

Retrieval performance (nDCG@10) by reranker across public (BEIR) and internal benchmarks. Internal benchmarks are grouped by type of retrieval documents: ‘Long Context’ targets retrieval of documents with several thousand words. ‘Parsed PDFs’ groups retrieval datasets with PDFs parsed across financial, healthcare, biomedical, healthcare domains, spanning 6 languages in total. ‘Semi-Structured’ evaluates retrieval on mixtures of structured data (e.g., metadata of products) and free-form text.

Pareto-efficient frontier of retrieval latency (ms for 30 x 500 tokens; time to process 15,000 tokens) versus average retrieval performance (nDCG@10) by reranker pooled over all benchmarks across languages and domains (cf. Plot 1, Plot 3, Plot 4)

To support different search workloads, Rerank 4 is available in two versions. Fast is a smaller model designed for use cases where both speed and accuracy are important. Pro, meanwhile, is optimized for tasks that require deeper reasoning, analysis and ironclad precision.

Rerank 4 Fast - where speed wins:

E-commerce. Prospective customers benefit from high relevance product recommendations as well as the option to explore online catalogues with semantic search. Here, faster results equal a better shopping experience and higher conversion rates.
Programming. During a project, engineers need to quickly and conveniently reference documentation - like design specs - without derailing their creativity or collaboration.
Customer service. Enterprise help desk agents need to efficiently triage incoming tickets to resolve cases, prevent backlogs, and keep their customers happy.

Rerank 4 Pro - where depth matters:

Finance. When generating risk models, analysts rely on troves of market reports, regulatory filings, and transaction histories. More retrieval time means richer data analysis and more accurate scenarios modeled.
Healthcare. Clinicians routinely review patient records and clinical trial reports to identify appropriate courses of treatment. Accuracy and completeness of information can have devastating consequences for patient welfare.
Manufacturing. Quality engineers consult technical manuals and production logs to diagnose deviations in highly complex processes that led to product defects. More informed decision-making helps productivity, safety, and end customer satisfaction.

Superior performance in enterprise-specific and multilingual tasks

Our customers need to know that their semantic search is not just performant, but works for the queries, workflows, languages, and data types that constitute their everyday work. This means the model has to understand the nuances of each user and the problem they're trying to solve.

Benchmarking against industry standards confirms that Rerank 4 outperforms all current competitive alternatives in a broad array of enterprise domains, such as finance, healthcare, manufacturing and others.

Retrieval Performance (nDCG@10) by domain: ‘Workplace Productivity’ tests retrieval of workplace communication (e.g., e-mails, threads). ‘Finance’, ‘Healthcare’, and ‘Manufacturing’ group retrieval performance over several public and in-house domain-specific datasets, respectively.

Rerank 4 also leads the pack in multilingual performance. Users can deploy Rerank 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’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.

Self Learning

Semantic search is an investment for your company. Like any human-executed task, semantic search will see greater returns once it ‘learns’ the means to do its job more efficiently. On designing Rerank 4, we asked ourselves: how can we bake this concept into our product offerings?

Rerank 4 is our first reranking model with self-learning capability. In partnership with Cohere, users can customize their model for certain use cases - such as those they encounter most frequently - without the need for further annotated data. This might involve stating preferences for particular content types, for use of prescribed language or terminology, or directing the model to specific document corpora.

An example: imagine a Consumer Loan Specialist at a commercial bank. Their job is to evaluate borrowing applications and guide customers through the process of acquiring a loan. Each case requires frequent references to internal documentation relating to, say, approval criteria, product details, regulatory standards, or even company policies. Fast retrieval is key for customer satisfaction - but accuracy can’t be compromised either. At least not within this narrow domain.

Rerank 4 moves towards overcoming this tradeoff in speed and quality through Self Learning. With automated adaptation to domain-specific use cases, the accuracy of Rerank 4 Fast iteratively improves to become competitive with much larger market alternatives. In many of our tests, Rerank 4 Fast began to converge and even surpass out-of-the-box Pro performance. Self-Learning Rerank will soon bring continuous RAG optimization to North, advancing precision and robustness in domain‑specific workflows.

Industry-wide performance (nDCG@10) comparison on the ViDoRe V3 benchmark, demonstrating domain-level retrieval accuracy for Rerank 4 Fast with and without Self Learning. ViDoRe V3 is a publicly available benchmark that contains a large collection of visually rich, multimodal enterprise documents such as reports, invoices, forms, and technical documentation.

Looking further, we also explored how Rerank 4’s self-learning capability performs on entirely new search domains. Using healthcare-focused datasets that mimic a clinician’s need to retrieve patient-specific information - not just expertise from a given medical discipline - we found that enabling Self Learning produced consistent, substantial gains. The result: a clear and significant boost in retrieval quality for Rerank 4 Fast, across the board.

Getting started

Rerank 4 is now available on Cohere’s Platform, Amazon SageMaker AI (Fast & Pro) and Microsoft Foundry, with additional platform support coming soon.

Click here for pricing options, and consult our docs for learn more about Rerank 4 setup.

Rerank 4 can also be deployed into any Virtual Private Cloud (VPC) or on-premise environment. To learn more about options for large enterprise deployments, please contact our sales team.

Introducing Rerank 4: Cohere’s most powerful reranker yet