Mistral's Agentic Search Reads Documents Like an Analyst

Picture a junior analyst asked whether a client's depreciation policy changed over three fiscal years. She doesn't run one search and copy the first paragraph she finds; she opens three annual reports, flips to the same note in each, and compares them line by line. That workflow, mundane as it sounds, is exactly what most enterprise AI search systems cannot do, and it is what Mistral shipped on August 20 under the name Agentic Search.

Five tools instead of one shot

The French lab's new retrieval layer hands the model a small toolkit modeled on how people work with files: search for documents, open one, navigate within it, read the content, and grep for patterns. The model runs these in a loop, inspecting what it finds, refining its query, and verifying sources before it commits to an answer. Classic retrieval-augmented generation (RAG) setups chop documents into chunks and pull whatever looks similar to the question in one pass; when the answer lives across two sections, or inside a table, they routinely miss. The agentic loop exists to close that gap, because the model can notice it hasn't found enough and go look again.

The vendor's numbers, and the caveat that goes with them

Mistral's own benchmarks are striking. On FinanceBench, built from 368 SEC filings, accuracy jumps from 26.7 percent to 86 percent, while worst-case latency drops from 255 to 154 seconds. On OfficeQA Pro, a test spanning 696 US Treasury bulletins, accuracy climbs 45.6 percentage points. Token consumption falls by up to a third, since the model discards irrelevant material early instead of stuffing it all into context. The standard caution applies: every one of these figures comes from Mistral's own evaluation, no independent replication exists yet, and pricing hasn't been announced.

The tool ships in Mistral's Search Toolkit and its Studio and Vibe libraries, and runs both in the cloud and on-premises. That last detail is the strategic one. Combined with the company's regional inference push to keep data inside Europe, Mistral is assembling a coherent offer for organizations that want capable AI without their documents leaving the building.

A question worth stealing for your own project

From the projects we see, the failure mode of document AI is rarely the language model; it is retrieval choking on table-heavy PDFs and cross-referencing contracts. So if you are scoping an internal document assistant this year, borrow the framing this release suggests: start by asking whether the system can navigate your documents, then argue about vector databases. The projects that get that order right are the ones that survive contact with real archives.

Sources: Mistral AI