Mistral Agentic Search, the iterative document search
- Maxime Hiez
- Mistral AI
- 29 Aug, 2026
Introduction
Mistral AI announced on August 20, 2026 Agentic Search, a document search layer meant to replace the classic RAG in the cases where it fails. The principle changes : instead of a retrieval in a single pass, the model chains several steps to find, open, browse and verify the information before answering.
The problem of one pass RAG
The classic RAG splits the documents into chunks, computes their similarity with the question, and passes the best results to the model. This works as long as the answer fits in one chunk.
It fails as soon as the answer requires cross-checking several pages, reading a table in its context, or comparing two documents. The model can neither refine its search, nor open a document to verify, nor go back. It answers with what it was given, including when it isn’t the right section.
Five tools to navigate
Agentic Search puts five tools at the disposal of the model :
- search : Locates the relevant documents in the corpus from the existing index
- open : Opens a specific document
- navigate : Moves to a page, a section or an area of the document
- read : Retrieves the content at the location reached
- grep : Searches for a pattern inside an open document
The model decides the sequence. It can refine an unsuccessful search, open the source to confirm a figure, or compare two documents before concluding. It is the same way of working as an analyst facing a pile of reports.
The measured results
Mistral publishes two evaluations, with unusually large gaps :
- FinanceBench : 368 SEC regulatory filings and 150 questions. Mistral Medium 3.5 goes from 26.7% to 86% of accuracy with the full loop, that is +59.3 points.
- OfficeQA Pro : 696 US Treasury bulletins and 133 questions. Accuracy goes from 6.3% to 51.9%, that is +45.6 points, with Mistral Medium 3.5 as with GLM-5.2.
The gain isn’t limited to accuracy. Latency at the 90th percentile drops by up to 39.6% and token consumption by up to a third. The point is counter-intuitive : chaining several calls costs less than a single call stuffed with useless chunks.
Deployment modes
Agentic Search is offered in several forms :
- Mistral Search Toolkit : Available in the cloud and self-hosted
- Studio and Vibe Libraries : Direct integration in the Mistral tools
- Search Starter App : Starter application to validate the behavior locally
The self-hosting option is the point to remember for the organizations with data residency constraints. It places the document corpus and the search loop within the perimeter of the organization, which is rarely possible with managed search services.
Conclusion
Agentic Search tackles the weak link of enterprise RAG deployments, the quality of the retrieval rather than that of the model. The gaps measured on financial and administrative documents, two dense corpora full of tables, are significant. For an organization that already attempted a RAG project abandoned for lack of reliability, the approach deserves a new try, all the more since the reduction in latency and tokens goes in the same direction as the quality.
Sources
Did you enjoy this post ? If you have any questions, comments or suggestions, please feel free to send me a message from the contact form.
Don’t forget to follow us and share this post.
