Mistral's New Search Tool Lets AI Dig Deep Into Your Company's Documents, Not Just Skim the Surface
Agentic Search replaces single-pass document lookup with a multi-step reading loop, tripling accuracy on financial filings in benchmark tests.
Key points
- Mistral AI's Agentic Search roughly tripled correct answers on financial filing questions, jumping from 26.7% to 86% on the FinanceBench test.
- On table-heavy, multi-document questions, accuracy rose by 45.6 percentage points, from 6.3% to 51.9%, on the OfficeQA Pro benchmark.
- The system reduces response lag by up to 39.6% and cuts the number of AI-generated tokens (units of text the model processes) by up to one third.
- Agentic Search is available now through Mistral's Search Toolkit and runs on-site or in the cloud without sending sensitive data outside a company's own systems.
Most AI search tools work like a student who skims the index of a textbook, grabs the first few paragraphs that look relevant, and writes an answer without ever opening the chapter. That works fine for simple questions. It fails badly when the answer is buried in a footnote, spread across three tables, or hidden on page 94 of a 150-page filing.
Mistral AI's new Agentic Search is designed to fix that.
What does Agentic Search actually do differently?
Instead of one quick lookup, it gives the AI a small set of tools that let it read documents the way a human analyst would: open a file, jump to a section, scan for a specific term, and verify what it finds before answering.
The five tools are search (find relevant documents), open (open a specific one), navigate (jump to a page or section), read (pull the text there), and grep (scan for a pattern inside an open file). Think of it as giving the AI a proper filing cabinet to rummage through, not just a sticky note with three bullet points.
The old approach, called RAG (short for retrieval-augmented generation, a technique where an AI pulls snippets from stored documents before answering), retrieves a fixed set of text chunks and asks the model to answer in one shot. If the right chunk is not in that first batch, the answer is wrong, and the AI has no way to go back and check.
Agentic Search loops. The model can decide it needs more context, open a new document, follow a reference, and try again.
How big is the accuracy improvement?
The gains are large enough to matter in real decisions.
On FinanceBench, a standard test using 368 SEC financial filings averaging 147 pages each, Agentic Search pushed correct answers from 26.7% to 86%. On OfficeQA Pro, which tests multi-document questions filled with tables, the score rose from 6.3% to 51.9%.
Mistral tested two models: its own Mistral Medium 3.5 and a third-party model, Z.ai GLM-5.2. Both showed similar gains, which suggests the improvement comes from the search method itself, not from any one AI model.
| Benchmark | Old accuracy | New accuracy | Gain |
|---|---|---|---|
| FinanceBench | 26.7% | 86.0% | +59.3 pts |
| OfficeQA Pro | 6.3% | 51.9% | +45.6 pts |
| Response lag (p90) | baseline | up to 39.6% faster | , |
| Token use | baseline | up to 33% lower | , |
Who is this actually for?
Companies that sit on large volumes of sensitive documents: law firms with contracts, hospitals with clinical records, finance teams with regulatory filings. The system runs inside a company's own infrastructure, so the documents never leave.
For employees, the practical effect is that an AI assistant should stop giving confidently wrong answers to questions like "What was our effective tax rate in Q3?" and start returning answers it can actually point to on a specific page.
Agentic Search is available now through the Mistral Search Toolkit. No fine-tuning of the AI model is needed, which means the system should get better automatically as underlying AI models improve.
Common questions
Do my company's documents get sent to Mistral's servers?
No. Mistral says Agentic Search runs on-premises or in your own cloud environment, so sensitive files stay inside your existing security boundary.
Does my IT team need to rebuild our document storage from scratch?
No. Agentic Search works on top of whatever document index a company already has in place, using the five navigation tools to go deeper into it.
