Mistral's New Agentic Search Lets AI Actually Read Your Documents
The French AI lab launched a retrieval system that lets models open, navigate and verify complex documents themselves, more than tripling accuracy on a finance benchmark.
Mistral, the French AI lab, launched a feature called Agentic Search on August 20. It’s a new way for AI models to dig through documents, tables, footnotes and all, before answering a question, instead of relying purely on the older method most AI tools still use.
That older method is called RAG, retrieval-augmented generation: the system searches a database for chunks of text that look relevant, hands them to the model, and hopes the answer is in there. Agentic Search gives the model five extra abilities instead, search, open, navigate, read and grep (a way of scanning text for specific patterns), letting it behave more like a person actually working through a document: opening a file, checking a table, following a reference to another section, refining its own query if the first attempt comes up short. On FinanceBench, a benchmark that tests question-answering across hundreds of real SEC financial filings, Mistral says accuracy jumped from 26.7 percent with older methods to 86 percent with Agentic Search. It’s available now through Mistral’s Search Toolkit for custom setups, and built directly into Libraries inside Mistral’s Studio and Vibe products.
What’s actually going on here: anyone who’s tried asking an AI chatbot a precise question about a long PDF, a contract, a financial report, a technical manual, knows the frustrating result: a confident-sounding answer that’s subtly wrong because the model only saw a fragment of the relevant text. That’s the real-world failure mode RAG-based tools run into constantly. Letting the model actively navigate a document, the way you’d flip to page 40 because that’s where the numbers you need actually live, closes a lot of that gap. It also reportedly does it more efficiently, fewer back-and-forth turns and less token usage (tokens are the chunks of text AI systems are billed by), which matters for anyone running this at scale. For Mistral, this is also a play to stay relevant in enterprise AI against much larger, better-funded US labs, by shipping practical tools that solve a specific, common pain point well.
What this means for you: if you or your company deal with long, dense documents, financial reports, legal contracts, technical specs, and use AI tools to search or summarize them, this is the kind of improvement that could meaningfully cut down on hallucinated or incomplete answers. For most casual chatbot users, it won’t be visible directly, but it’s a good sign of where AI search tools generally are heading: less “guess from a fragment,” more “actually go look it up properly.”
Sources
XPeng's Robot Unit Just Raised $900 Million for Its Humanoid Push
XPeng's robotics subsidiary, backed by Alibaba and Tencent, is now valued at over $6.3 billion ahead of mass production of its IRON humanoid robot.
