Agentic Search. More accurate and efficient results from your AI systems.
Mistral AI
Mistral launched Agentic Search, a tool that can open and check documents instead of just grabbing snippets. It says that cuts errors, tokens, and latency on hard financial and office docs.
Based on reporting by Mistral AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Mistral AI is pushing a more hands-on style of search for enterprise systems. Its new Agentic Search layer is meant to do more than pull back a few text chunks and hope for the best. It can search, open, navigate, read, and grep through documents before answering, and Mistral says that makes a real difference on dense material like filings, contracts, manuals, and scanned PDFs.
The pitch is simple: one-shot retrieval often breaks when the answer is buried in a table, a footnote, or a clause that lives several pages away from the first hit. Agentic Search adds a multi-step loop so the model can refine its search, inspect a document, move to a section, and verify what it found. Mistral says this works through its Search Toolkit and through Libraries in Studio and Vibe, and that it can run across cloud or on-premises setups without crossing isolation boundaries.
The company backs that with two benchmarks. On FinanceBench, which covers 368 SEC filings and 150 questions, Mistral says the system lifts accuracy from 26.7% to 86% on financial filings, and on OfficeQA Pro, a harder set built from 696 Treasury Bulletins and 133 questions, it reports a jump from 6.3% to 51.9%. The same setup also cuts p90 latency by up to 39.6% and reduces token use by up to one-third.
Mistral is careful to frame this as a floor, not a ceiling. The tests used default chunking and default ranking in the out-of-the-box toolkit, with no tuning. The company also says the results were consistent across Mistral Medium 3.5 and Z.ai GLM-5.2, which is the real story here: better retrieval now depends less on a perfect chunking trick and more on letting the model keep looking.
It’s a neat answer to a very old enterprise problem. The data people actually care about is usually the most annoying data to search. Mistral’s bet is that the model should stop acting like a one-shot autocomplete machine and start behaving more like a careful analyst with a filing cabinet and a flashlight.
My take — AI-written commentary, not fact-checked reporting
This is the right direction, and also a quiet admission that chunk-and-pray RAG was never the whole answer. The industry loves selling “AI over your data” while ignoring the fact that the data is often buried, messy, and allergic to neat snippets. Open tools that can actually inspect sources beat glossy demos every time.
Read more about this at: Mistral AI