--- title: "Mistral's Agentic Search lets models open and read the document" url: https://blog.vomini.com/blog/mistral-agentic-search-documents date: 2026-08-21T13:40:00+00:00 updated: 2026-10-27T22:42:00.046738+00:00 author: "Roberto R." section: "News" language: en site: "Vomini" license: free to quote with attribution --- # Mistral's Agentic Search lets models open and read the document > Mistral released Agentic Search, a set of five tools that let a model search, open and read inside long documents before it answers. On a test of questions about SEC filings, Mistral reports accuracy going from 26.7%… On 20 August, Mistral released Agentic Search, a way for AI models to look things up inside a company's documents in several steps. It is aimed at the files where answers hide in a table on page 90: financial filings, contracts, manuals, government records. Most company chatbots today use RAG, short for retrieval-augmented generation. The system grabs a handful of text snippets that look relevant and asks the model to answer from those snippets in one pass. When the answer sits in a footnote the snippets missed, the model has no way to go back and look. ## Five tools instead of one pass Agentic Search gives the model five tools that work on top of an existing search index: `search`, `open`, `navigate`, `read` and `grep` (a command that finds a word or pattern inside an open file). The model can search, open a promising file, jump to a section, read it and search again if the first results were weak. Mistral says this needs no fine-tuning, meaning no extra training of the model on your data. It is available through the Mistral Search Toolkit and is built into Libraries in Mistral's Studio and Vibe products. ## What Mistral measured Mistral ran its tests with default settings and two models, its own Mistral Medium 3.5 and Z.ai's GLM-5.2. On FinanceBench, a set of 150 questions over 368 SEC filings that average about 147 pages each, the best result went from 26.7% correct with one-pass RAG to 86% with Agentic Search. On OfficeQA Pro, which asks table-heavy questions across several documents, accuracy went from 6.3% to 51.9%. Mistral also reports slow-case response times falling by up to 39.6% and token use falling by up to a third, because the model repeats fewer broad searches. These are Mistral's own numbers on public benchmarks. The post does not list prices for the new tools. ## When plain search is enough Mistral is upfront that the old approach still fits some jobs. For short, clean documents, high-volume keyword lookups and questions where you already know where the answer lives, one-pass retrieval is usually enough. Agentic Search is for the cases where the model has to dig, compare sources or check a number against the page it came from. ## What to do with this If your team has an internal assistant that answers questions about long PDFs, collect ten questions it got wrong and note where the right answer actually was. If most of them sit in tables or footnotes, a multi-step search like this one is worth a trial. If they are simple lookups that failed, the problem is more likely the index itself. ## What we think The biggest gain in Mistral's tests came from letting the model search more than once, before any of the new reading tools were added. That matches what anyone who has watched a chatbot miss an answer in a long PDF would expect. The tradeoff is time and tokens per question, so we would keep it for the questions that need it. ## Sources - [Mistral AI](https://mistral.ai/news/agentic-search/), 20 Aug 2026 --- Originally published at https://blog.vomini.com/blog/mistral-agentic-search-documents