Mistral introduces Agentic Search for complex document retrieval

A multi-step retrieval layer

Mistral AI introduced Agentic Search, a retrieval layer designed to let AI systems search, navigate and verify information within complex documents and across multiple data sources. The company says the system uses a multi-step retrieval loop rather than requiring a model to answer from a fixed set of chunks selected in a single search.

Agentic Search is available through Mistral Search Toolkit and is built into Libraries in both Studio and Vibe. Mistral says its portable, open tooling can work with sensitive, domain-specific data in the cloud or on-premises without moving that data across isolation boundaries.

Five tools for document navigation

Agentic Search builds on an organization’s existing search index and gives a model five tools:

• search finds relevant documents through the index.
• open opens a selected document.
• navigate moves to a page, section or region.
• read retrieves content from that location.
• grep finds a pattern within an open document.

Mistral says a model can use these tools to inspect initial results, revise its search, move to relevant sections and read source material before producing an answer. The company says the tools require neither fine-tuning nor model-specific training.

The Mistral Search Toolkit also includes open modules for ingesting, embedding and indexing data. The company identifies long documents, questions spanning multiple sources, answers requiring verification, tables and structured documents as intended uses for Agentic Search. It says conventional indexed retrieval can remain sufficient for direct lookups, high-volume searches and simple questions with predictable source locations.

Benchmark results reported by Mistral

Mistral tested Agentic Search with the default Mistral Search Toolkit configuration, including default chunking and ranking without tuning. The tests used Mistral Medium 3.5 and Z.ai GLM-5.2.

On FinanceBench, which covers 150 questions drawn from 368 SEC filings, Mistral reports that correctness increased from 26.7% to 86%. The company says moving from one-shot retrieval to an iterative search-only loop raised accuracy by 47.3 percentage points for Mistral Medium 3.5 and 52.6 points for GLM-5.2.

Mistral reports that adding the open, navigate, read and grep tools produced further accuracy gains of 8.7 percentage points for Mistral Medium 3.5 and 6.7 points for GLM-5.2. According to the company, the full loop used 23.9% fewer tokens than the search-only loop with Mistral Medium 3.5 and 33.7% fewer with GLM-5.2.

For OfficeQA Pro questions involving tables and multiple documents, Mistral says results increased by 45.6 percentage points, from 6.3% to 51.9%. The company also reports reductions of up to 39.6% in p90 latency and up to one-third in token consumption.

Source: Mistral AI’s “Introducing Agentic Search” announcement, published August 20, 2026.