Mistral Search Lifts Document Accuracy 27% to 86%

The Benchmark Mistral Chose to Publish

Mistral built Agentic Search around a specific test: FinanceBench, a benchmark of 368 real SEC filings and 150 questions that each require finding and cross-checking a specific fact buried inside dense financial disclosures.

Under Mistral's own one-shot retrieval-augmented generation baseline, where a model reads whatever a single search pass returns and answers from that alone, FinanceBench accuracy sat at 26.7 percent. Agentic Search, which lets the model run repeated rounds of searching, opening individual filings, and verifying its own findings before it answers, raised that figure to 86 percent, according to Mistral's own published results. A benchmark score moving from roughly one correct answer in four to better than eight in ten is a significant result for a single product update focused on the retrieval layer alone.

The Failure Mode Agentic Search Targets

One-shot RAG fails enterprise users in a specific, repeatable way: it retrieves a handful of document chunks in a single pass and forces the model to answer from only those chunks, even when the real answer sits in a table, a footnote, or a different filing entirely.

Mistral built Agentic Search around five operations that resemble ordinary file-system commands: search, open, navigate, read, and grep. Instead of guessing from a fixed set of retrieved snippets, the model can open a specific document, navigate to a table, read the surrounding context, and search again if the first answer looks incomplete. For a question like finding a single indemnification clause across hundreds of SEC filings, that iterative process is closer to how a trained analyst actually works than a single database lookup ever was. Mistral frames this as evidence that the retrieval step has been the primary constraint in enterprise AI search.

A Second Benchmark Rules Out a Fluke

Mistral did not stop at one benchmark before publishing these accuracy numbers.

OfficeQA Pro is a separate test built from 696 Treasury bulletins and 133 questions, covering a different document type and a different kind of question than FinanceBench. Accuracy on OfficeQA Pro rose from 6.3 percent under one-shot retrieval to 51.9 percent with the full agentic loop, Mistral said. The absolute numbers are lower than FinanceBench because OfficeQA Pro questions are harder by design, and the shape of the result is the same: a one-shot retrieval baseline that answers barely one question in twenty correctly, and a multi-step search-and-verify loop that recovers most of the gap.

What the Table Shows About Cost, Not Just Accuracy

Multi-step search sounds like it should cost more, and the opposite turned out to be true in Mistral's own testing.

Two benchmarks, one latency measurement, and one token measurement all moved in the same direction, which is the detail Mistral is highlighting for enterprise buyers weighing whether a more thorough search process is worth the added engineering effort.

What On-Premises Actually Covers for European Buyers

Agentic Search runs in Mistral's cloud or fully on a customer's own infrastructure, and that on-premises option is the detail most relevant to EU and UK enterprises evaluating AI search vendors under data-residency requirements.

Most AI products that combine large language models with document retrieval are built and hosted exclusively on US hyperscaler infrastructure, which requires European buyers doing vendor due diligence to accept that sensitive documents leave the organization's own environment. Mistral, a French AI lab, is offering an alternative deployment path where the retrieval and search infrastructure can run inside a customer's own data center or private cloud. That is a real and useful distinction, and it is narrower than full data sovereignty: an on-premises deployment option describes where the software runs, not an independent audit, a specific compliance certification, or a guarantee about how the underlying model was trained. Enterprises should treat it as one input into a vendor evaluation, alongside their own compliance review.

Read next: 4,000 Apps in 30 Days: Cloudflare's AI Wager  |  Mistral Just Split Sovereign AI Into Two Bets