Cohere’s Rerank 3.5 improves the relevancy of search results by providing systems with a boost in semantic understanding of complex business data. While this model is generally useful, business tend to use it to: 1. Enhance Agentic AI and Retrieval-Augmented Generation (RAG) Systems: Before delivering a useful answer to business questions, AI systems often need to sift through large amounts of data to gain relevant context. This is costly and time consuming. Rerank 3.5 makes this process more efficient, rapidly and precisely searching potentially relevant documents so that only the most relevant are passed to generative AI models. This leads to faster and more accurate responses at a lower total cost of ownership. 2. Improve Enterprise Search Systems
Employees use enterprise search systems to quickly locate relevant information within their business. With just a few lines of code, and with minimal latency cost, Rerank 3.5 makes these existing systems more intelligent. It provides a boost in semantic understanding and an ability to search data formats which are traditionally inaccessible (e.g. tables, code, multilingual, JSON) such that users get more relevant answers. Customers find this model useful in scenarios with strict latency, throughput, and accuracy requirements. Rerank 3.5 works across common business languages and data formats. It excels at answering complex queries which require reasoning.

Evaluation

To understand Cohere’s Rerank 3.5 performance, we compare key capabilities with a set of alternative approaches over a variety of robust internal evaluations. Search systems often fail to retrieve relevant information when users implicitly or explicitly express constraints on what they would like returned. We identified that this was partially due to traditional systems lacking the ability to reason. Rerank 3.5 shows substantial improvements in this area, understanding complex multifaceted questions that other search systems fail to answer.

Model	Retrieval Accuracy on Data Requiring Reasoning
BM25	43.53%
Dense Embeddings	50.64%
Hybrid Search	48.80%
Cohere Rerank 3	27.91%
Cohere Rerank 3.5	81.59%

Reasoning Datasets are adversarial datasets where the user bounds a semantic search with implicit and explicit criteria. Reasoning dataset is measured as P@1 out of 2. This capability is particularly helpful for businesses operating within specialized industries such as finance, government, energy, manufacturing, and healthcare. For example, on a financial services dataset we curated to be generally representative for common use cases, Rerank 3.5 performance was +23.4% better than Hybrid Search and +30.8% better than BM25. We expect organizations in these industries to observe similar improvements when evaluating performance on their data. Rerank 3.5 also offers industry-leading multilingual capabilities. It can search across data in 100+ languages, with state-of-the-art accuracy on the following 10 global business languages: Arabic, Chinese, French, German, Hindi, Japanese, Korean, Portuguese, Russian, and Spanish.

Model	Retrieval Accuracy on Multilingual Data
BM25	38.19%
Dense Embeddings	53.83%
Hybrid Search	52.10%
Cohere Rerank 3	52.27%
Cohere Rerank 3.5	62.18%

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