Mistral has launched Agentic Search, a new retrieval tool designed to let AI models navigate and verify information inside complex documents rather than relying on single-pass search.
The tool adds a multi-step retrieval loop, giving models five capabilities: search, open, navigate, read and grep, allowing them to inspect documents, refine queries and follow references before answering, rather than working from a fixed set of retrieved text chunks. It is available through Mistral's Search Toolkit for custom deployments, and built into Libraries within Mistral's Studio and Vibe products for out-of-the-box use.
Mistral claims the approach delivers significant accuracy gains on two benchmarks. On FinanceBench, which tests question-answering across hundreds of SEC filings, the company says accuracy rose from 26.7 percent to 86 percent. On OfficeQA Pro, a benchmark built around scanned Treasury Bulletins, Mistral reports a 45.6 percentage point improvement, taking accuracy from 6.3 percent to 51.9 percent.
The company also claims the tool reduces computational overhead, cutting p90 latency by as much as 39.6 percent and lowering token consumption by up to a third, as targeted navigation replaces repeated broad searches.
Mistral positions Agentic Search as most useful for long, dense or multi-source documents such as filings, contracts and technical manuals, while suggesting simpler indexed retrieval remains sufficient for short documents or straightforward lookups.
