French company Mistral has released Agentic Search, a tool that helps artificial intelligence search corporate documents much like a human would. Existing search engines scanned an archive only once, extracted a few snippets, and the model had to answer solely based on them. The new tool, however, can search repeatedly, open a specific document, navigate to a selected page or table, and examine it. In a test based on US securities filings, this increased answer accuracy from just under 27 percent to 86.
Where one-time searches fall short
Today’s AI-powered enterprise search generally works in a simple way. The system finds several seemingly relevant passages of text and presents them to the model along with the question. If the answer is actually contained in them, the result is correct. But if the required information is buried in an overlooked table or an entirely different document, the model cannot access it. It answers only from the original source materials and cannot request another report or additional context.
According to Mistral, this approach has three weaknesses. The model cannot decide on its own that the initial results are insufficient. It also cannot open a more complex document and find the correct table within it, even if the system has correctly identified the document. It also lacks the ability to make a second attempt, verify a number it has found, or compare two different sources.
Five simple commands
Agentic Search builds on a company’s existing index and adds five tools with clear names: search finds documents across the entire archive, open opens a specific one, navigate moves the model to a page or section, read extracts the content from that location, and grep searches the open document for a specified pattern.
Mistral demonstrates this with an example from US government bulletins. A question asking for the total monthly defense expenditure for 1953 stopped halfway when using a one-time search because the documents found covered only part of the year. With the new approach, the model used a second, rephrased query to find a bulletin from February 1954, opened page fifteen, examined a table containing all twelve months, and calculated the result.
An important detail is that these tools do not require additional model training or modifications for individual model families. According to Mistral, search quality will therefore improve as models become better at reasoning and using tools, without the company having to make changes to its infrastructure.
What the new approach is suitable for and when it is unnecessary
The company specifies where deploying the new tool makes sense. This includes long documents such as contracts, reports, manuals, or technical specifications, as well as queries requiring comparisons across multiple sources, answers that must cite a specific location in the text, and, finally, tables or scanned files whose meaning depends on rows, columns, and surrounding context.
Conversely, for short and clearly structured documents where the answer is almost certainly contained in the first snippet found, for broad passage searches, or for queries where the exact location of the answer is known, a conventional index remains the better choice. Paying for unnecessary document traversal only adds costs without providing any benefit.
Even with complex documents, however, human responsibility remains unchanged. An accuracy rate of 86 percent means that roughly one in seven answers is incorrect. If the result is intended for audit documentation or an official filing, it should still be treated only as a good first draft, not as finished work.
Where it runs and what it costs
Agentic Search is available through the Mistral Search Toolkit for teams building their own search pipelines, and through the Libraries feature integrated into the Studio and Vibe products, where the tool can be used directly. The quickest way to try it is with the starter application on GitHub, which creates a local index from a user’s own document archive using the default settings.
It can run in the cloud or entirely on a customer’s own servers, which represents a crucial difference for European and British companies with strict data residency requirements compared with offerings built exclusively on US infrastructure. However, this is a description of the technical setup, not an independent audit or a specific certification, so companies must still conduct their own regulatory compliance assessments. According to the price list, indexing with the Libraries feature costs one dollar per million tokens, individual queries cost one cent, document text recognition costs three dollars per thousand pages, and the Vibe Pro plan costs just under fifteen dollars per month.
