Mistral announced Agentic Search on 2026-08-20, a retrieval layer that replaces one-shot retrieval-augmented generation with an iterative loop. It is delivered through the Mistral Search Toolkit and built into Libraries in both Studio and Vibe, and it operates over an organization’s existing search index rather than requiring a new one. The model is given five tools - search, open, navigate, read and grep - and can decide to fetch a different document, jump to a specific table, follow a reference or verify a claim before answering.
Mistral’s argument against conventional RAG is stated in three parts: retrieval without reasoning, because the model is stuck with whatever chunks the first pass returned; a chunk-level ceiling, because critical facts sit in a particular table, footnote or clause inside a long multimodal document; and no iteration, because one pass cannot refine a query or compare sources. The example the company uses - asking for a company’s effective tax rate in Q3, where an index finds the right filing but cannot open it and read the surrounding context - is a fair characterization of where document QA actually fails in practice.
The reported numbers are large and are Mistral’s own. The company says correctness on financial filings goes from 26.7 percent to 86 percent on FinanceBench, roughly a 3x gain, and that on the table-heavy multi-document questions of OfficeQA Pro it measures a 45.6 point improvement, from 6.3 percent to 51.9 percent. It also claims p90 latency falls by up to 39.6 percent and token consumption by up to a third, on the reasoning that targeted navigation avoids repeated broad searches. A jump from 6.3 to 51.9 percent says as much about how badly the baseline performs on that task class as about the new system, and none of it has been independently reproduced.
The sovereignty angle is the part a European buyer will care about. Mistral emphasizes that the tooling is portable and open so that customers can “unlock value from your data without crossing your isolation boundaries in the cloud or on-premises” - consistent with the in-region inference and European compute push the company announced nine days earlier. For anyone who has run a disappointing enterprise RAG pilot, the honest framing of this release is that the retrieval loop, not the model, was frequently the binding constraint, and that fixing it is now a product rather than a research project. The efficiency claims deserve a pilot on your own corpus before they are believed.