Mistral launches Agentic Search for RAG

Mistral launched Agentic Search on August 20, a retrieval system built around five tools, search, open, navigate, read, and grep, that let a model iteratively dig through long or dense documents instead of relying on a single retrieval pass. On FinanceBench, a set of 368 SEC filings averaging about 147 pages each, adding just the search loop roughly tripled accuracy, and the full navigation toolchain pushed it to 86%, up from 26.7% for one-shot retrieval. On OfficeQA Pro, a set of 696 scanned, table-heavy government financial PDFs, the full agentic loop lifted accuracy by 45.6 points to 51.9%.

The system also cut latency: FinanceBench p90 response time dropped from 255 seconds to 154 seconds, and mean latency fell from 108 to 71 seconds, with token usage down by up to a third. It works alongside existing search indexes rather than replacing them, and deploys on-premises or in the cloud.

What it means for you

Agentic Search is live now through the Mistral Search Toolkit for custom integration, or built into Studio and Vibe if you’re already on those. The accuracy gains are concentrated exactly where one-shot RAG tends to fail, long filings and table-heavy PDFs, so if your pipeline is built for short, clean documents, this is a narrower fit; if you’re pulling numbers out of financial filings or scanned forms, it’s worth a direct comparison against whatever retrieval setup you’re running now. It’s a similar bet to what Ramp’s model router makes on the model-selection side: treat retrieval and routing as places to add a decision loop rather than a single fixed step, and pair it with disciplined MCP connector governance if you’re wiring it into tools that touch sensitive documents.