Mistral has launched Agentic Search, a retrieval layer designed to enhance the accuracy and efficiency of AI systems. This technology enables models to navigate, read, and verify information within complex documents, improving search results for enterprises. Agentic Search is available through Mistral Search Toolkit and Libraries, offering support for sensitive data and improved access to existing indexes. The system introduces a multi-step retrieval loop for finding, inspecting, and verifying information across data sources, wherever they are stored. Source: mistral
Agentic Search delivers significant improvements in search accuracy, achieving up to 3x correctness on financial filings, from 26.7% to 86%, based on FinanceBench. On table-heavy, multi-doc questions from the OfficeQA Pro benchmark, it measures a +45.6 point gain, increasing accuracy from 6.3% to 51.9%. The system also reduces latency and token use, with targeted navigation enabling Agentic Search to cut p90 latency up to 39.6%. Fewer repeated searches reduce token consumption by up to one-third. Source: mistral
Traditional retrieval methods, such as one-shot RAG, often fall short when dealing with dense and complex data. These methods retrieve a fixed set of text chunks and require models to answer in a single pass, limiting their ability to navigate long reports, follow references, or verify underlying evidence. Agentic Search addresses these limitations by allowing models to inspect findings, refine searches, and access specific sections of documents before providing answers. This approach ensures more reliable results for questions that require cross-referencing multiple sources or verifying specific data points. Source: mistral
