Agentic Search: More Accurate and Efficient Results in AI Systems

Mistral has launched Agentic Search, a retrieval layer capable of navigating, reading, and verifying information within complex documents. This tool utilizes five tools—search, open, navigate, read, and grep—based on existing search indexes to perform multi-step information retrieval.

Traditional RAG (Retrieval-Augmented Generation) approaches retrieve fixed text chunks all at once to generate answers, leading to reduced accuracy when information is scattered across long reports or multiple documents. In contrast, Agentic Search enables models to perform iterative retrieval by opening documents as needed, navigating to specific sections, and cross-verifying sources.

In actual performance evaluations, Agentic Search increased the accuracy on FinanceBench from 26.7% to 86%, a more than threefold improvement. Additionally, on the OfficeQA Pro benchmark, accuracy for table-centric complex document questions rose from 6.3% to 51.9%, an increase of 45.6 percentage points.

Improvements were also observed in terms of efficiency. Through goal-oriented retrieval, p90 latency was reduced by up to 39.6%, and token consumption was cut by up to one-third by reducing repetitive searches. This feature is available via the Mistral Search Toolkit and Libraries, allowing secure handling of sensitive data in both cloud and on-premises environments.