Mistral Introduces Navigable Search Loops to Replace One-Shot Document Retrieval

Mistral Agentic Search is a search layer designed to overcome the limitations of traditional one-shot RAG methods. This system provides a multi-step retrieval loop that allows AI models to navigate, read, and verify information within complex documents. This enables more accurate results even in dense documents such as financial reports or legal contracts.

Traditional RAG methods generate answers based solely on fixed text chunks retrieved during the initial search, leading to performance degradation on complex questions that require following references within a document or comparing multiple sources. In contrast, Agentic Search utilizes five tools—search, open, navigate, read, and grep—to allow the model to directly navigate and verify the necessary information.

Benchmark test results show that Agentic Search improved accuracy on FinanceBench from 26.7% to 86%, a more than threefold increase, and on OfficeQA Pro from 6.3% to 51.9%, a rise of 45.6 percentage points. Additionally, it demonstrated efficiency by reducing p90 latency by up to 39.6% through targeted navigation and minimizing token usage by up to 1/3 by reducing redundant searches. This feature is available via the Mistral Search Toolkit and Studio, as well as Libraries within Vibe.