Mistral launches Agentic Search, a multi-tool retrieval system that triples accuracy on financial documents

Mistral AI introduced Agentic Search on August 20, 2026, a retrieval system that lets its models iteratively search, inspect, and verify information across an organization's documents instead of relying on a single retrieval pass.

What's new

Mistral describes the product directly: "Mistral Agentic Search helps enterprises get better results from their AI systems by letting models search and navigate their organization's most complex data and documents." Rather than a one-shot lookup, "Agentic Search introduces a multi-step retrieval loop for finding, inspecting, and verifying information across data sources, wherever it is stored," giving models five tools — search, open, navigate, read, and grep — to refine a query across rounds.

The benchmark numbers are the headline. On FinanceBench, a set of 368 SEC filings with 150 questions, Mistral reports "3x correctness on financial filings, from 26.7% to 86%," with p90 latency dropping from 255 seconds to 154 seconds and token usage falling roughly 24-34% depending on the underlying model. On OfficeQA Pro, built from 696 Treasury Bulletins with 133 questions, accuracy rose by a "+45.6 point gain (6.3% to 51.9%)," with the navigate tool alone improving accuracy by "up to 35.6%." Mistral also cites latency reductions of "up to 39.6%" and token-consumption cuts of up to a third across its test suite.

Agentic Search is shipping now through the Mistral Search Toolkit and is built directly into Studio and Vibe, with support for both cloud and on-premises deployment for sensitive, domain-specific data. Mistral points developers to a "Search Starter App" for local indexing as an entry point.

Context

Retrieval-augmented generation has largely relied on single-pass embedding search, which struggles on dense, structured documents like SEC filings and government bulletins where the right answer is buried in tables or cross-references rather than a single passage. Mistral's framing — search as an iterative, tool-using loop rather than a one-shot lookup — follows a broader industry shift toward agentic patterns for tasks that used to be handled by static pipelines, mirroring how coding and browsing tools have moved from single-call retrieval to multi-step agent loops over the past year.

Why it matters

The FinanceBench and OfficeQA Pro results are large enough to change deployment decisions for enterprises doing document-heavy AI work: a jump from 26.7% to 86% correctness on financial filings is the difference between a tool that needs constant human verification and one that can be trusted for first-pass analysis. Cutting token usage and latency at the same time as improving accuracy also addresses the usual agentic-AI tradeoff, where multi-step reasoning tends to cost more and run slower even when it's more accurate. Availability on-premises is notable too — it targets exactly the regulated, sensitive-data customers (finance, government, healthcare) who have been the most hesitant to adopt cloud-only agentic retrieval.