Introducing Agentic Search

Mistral AI has launched Agentic Search, a retrieval layer designed to enhance AI systems' ability to navigate, read, and verify information within complex documents. Available through the Mistral Search Toolkit and Libraries, it supports sensitive domain-specific data while maintaining cloud or on-premises isolation boundaries. The tool introduces a multi-step retrieval loop for finding, inspecting, and verifying information across diverse data sources. Mistral AI reports that Agentic Search improves search results by up to 3x correctness on financial filings and achieves a +45.6 point gain on OfficeQA Pro benchmarks for table-heavy, multi-document queries.

Agentic Search addresses limitations of traditional one-shot retrieval-augmented generation (RAG) methods, which often fail to navigate long reports, follow references, or verify evidence. Unlike conventional approaches, Agentic Search enables models to refine searches, open documents, navigate to specific sections, and read source material before responding. This iterative process reduces the need for repeated searches, lowering token consumption by up to one-third and decreasing p90 latency by up to 39.6%. The tool operates without requiring fine-tuning or model-specific training, allowing retrieval quality to scale with model capability.

The tool provides five retrieval tools—search, open, navigate, read, and grep—resembling familiar file-system operations. These tools allow models to identify likely sources and inspect relevant sections, such as tables, clauses, or footnotes, within long documents like financial filings or legal contracts. Mistral AI emphasizes that Agentic Search is designed for proprietary and confidential data, enabling secure access to sensitive materials without compromising isolation boundaries. The tool is integrated into Mistral Search Toolkit, Libraries in Studio, and Vibe, supporting both cloud and on-premises deployments.

Mistral AI highlights that Agentic Search is particularly effective for questions requiring multi-step reasoning, such as calculating the total sum of U.S. national defense expenditures for 1953 using monthly data. Traditional RAG methods would struggle to locate and aggregate such dispersed information, whereas Agentic Search can refine searches, open relevant documents, and extract precise values. The company positions Agentic Search as a solution for enterprises seeking to leverage AI for complex data analysis while maintaining data security and compliance.