Mistral released Agentic Search on August 20, 2026, a retrieval layer that lets an AI model hunt through documents on its own. Instead of answering from a single pass of search results, the model can search again, open a file, move to the page it needs, and check the source before responding. On a financial document benchmark, accuracy climbed from 26.7 percent to 86 percent.
Not Stopping at One Retrieval
Traditional RAG queries an index once, hands the top-ranked text chunks to the model, and asks for an answer. That works when the answer sits in one of those chunks. It breaks down with a few-hundred-page contract or an annual filing, where the figure you need is buried in a specific table, footnote, or clause. The model has to answer from whatever it was handed, and it cannot go looking elsewhere when it notices something is missing.
Agentic Search turns that single retrieval into a loop. The model searches, inspects the results, searches again if they fall short, opens a promising document, moves around inside it, and reads the relevant passage, repeating as needed. The index still handles the job of narrowing down which documents look relevant, while the model decides where to dig and how deep.
Five Tools Handed to the Model
Agentic Search gives the model five capabilities that resemble ordinary file operations: search finds relevant documents through the existing index, open pulls up a specific document, navigate moves to a page, section, or region within it, read retrieves the content at that location, and grep looks for a word or pattern inside an open document.
Combined, these five let the model search, weigh the results, drill in, and reformulate its question when necessary. Chunks it has already seen can be excluded from later retrievals, so it does not keep circling back to the same hits.
Mistral says these tools require no fine-tuning and no model-specific training. They behave the same way with Mistral models and third-party models alike, which means retrieval quality improves as models get better at reasoning, without touching the indexing setup. That is a direct answer to a familiar frustration, where a chunking strategy quietly caps how accurate retrieval can ever get.
Results Across 53,900 Pages of SEC Filings
Mistral ran its tests with two models of different sizes, Mistral Medium 3.5 and Z.ai's GLM-5.2, using the Search Toolkit in its default configuration with no benchmark-specific tuning.
The first benchmark is FinanceBench, covering 150 questions over 368 SEC filings, 10-K, 10-Q, and 8-K documents averaging 147 pages each and totaling roughly 53,900 pages. Simply moving from one-shot RAG to a loop where the model can repeat its own searches raised accuracy by 47.3 points with Mistral Medium 3.5 and 52.6 points with GLM-5.2, roughly a threefold improvement in both cases.
Adding the navigation tools, open, navigate, read, and grep, pushed accuracy up another 8.7 points with Mistral Medium 3.5 and 6.7 points with GLM-5.2. Overall, Mistral reports a move from 26.7 percent to 86 percent.
Better Answers on Fewer Tokens
What makes this notable is that the accuracy gain does not come with a matching cost increase. The full loop with navigation consumed fewer tokens than the search-only loop, down 23.9 percent with Mistral Medium 3.5 and 33.7 percent with GLM-5.2, because repeated broad searches get replaced by targeted moves.
Latency follows the same pattern. On FinanceBench, p90 response time dropped from 255 seconds to 154 seconds and mean latency from 108 seconds to 71 seconds. Mistral says targeted navigation cuts p90 latency by as much as 39.6 percent.
Reading Tables in Scanned PDFs
The second benchmark is OfficeQA Pro, built from 696 historical U.S. Treasury Bulletins spanning roughly 89,000 pages of scanned, table-heavy PDFs. Mistral used the 133-question pro subset, where answers are numeric and verifiable.
GLM-5.2 went from 6.3 percent to 51.9 percent, a 45.6 point gain, while Mistral Medium 3.5 improved by 27.1 points. Navigation contributed another 7.5 points for Mistral Medium 3.5 and 8.3 points for GLM-5.2, and cut the number of turns by up to 7.0 percent.
Mistral also points out that the surrounding tooling changes the outcome for the same model. Kimi's research puts GLM-5.2 at 41.4 percent on OfficeQA Pro, 10.5 points below the 51.9 percent Mistral measured on its own harness. The argument is that raw model capability is only part of the picture, and how the retrieval stack is assembled matters as well.
These figures all come from Mistral's own testing rather than independent evaluation, which is worth keeping in mind when reading them.
Running It Inside Your Own Boundary
Agentic Search is used through the Mistral Search Toolkit, which bundles the components for ingestion, embedding, indexing, and retrieval. Pipelines can be assembled by choosing parsers, chunking strategies, embedding models, and extractors, while indexing and ranking are tuned through Vespa schemas and relevance profiles. Query rewriting, reranking, and hybrid retrieval can be layered on as well.
The search tools can be exposed to agents over MCP, the Model Context Protocol. The index runs either self-hosted with Vespa and Docker or on a managed backend, so sensitive data can stay inside an isolated cloud environment or on-premises, which is the main selling point.
For teams that would rather not build a pipeline, Agentic Search is already built into Libraries in Mistral Studio and Vibe. The quickest way to experiment is the Search Starter App published on GitHub, which builds a local index with a default configuration so you can watch Agentic Search work on your own corpus.
Summary
Agentic Search proposes moving retrieval out of the preprocessing step and into the work the model itself performs. Letting the model repeat its searches alone brought accuracy close to three times higher, and adding tools for moving inside documents raised accuracy further while reducing tokens and wait time. The numbers are Mistral's own measurements, but for anyone whose internal AI is struggling with long documents, they make a solid case that reworking the retrieval layer may pay off faster than reaching for a bigger model.
