Release2026-08-20
Mistral released Agentic Search, a retrieval layer that lets a model run several search steps and open, move through and read inside documents rather than answering from one batch of extracted text. It is available through the Mistral Search Toolkit and built into Libraries in Studio and Vibe, and can run in the cloud or on a company's own servers. Mistral reports large accuracy gains on two document benchmarks using its own and a third-party model, with no tuning applied.
What changed
Retrieval used one-shot chunk lookup, answering only from the first set of retrieved passages.
What it unlocks
Letting a model open a document, jump to a page or table and read it before answering, on an existing search index.
- 26.7% to 86% on FinanceBench
- 6.3% to 51.9% on OfficeQA Pro
- p90 latency 255s to 154s
What you need to act on it
- an existing search index built with Mistral Search Toolkit, or use of Libraries in Studio or Vibe
- developer setup for the toolkit route
Sources