On 20 August Mistral announced Agentic Search: instead of a one-off retrieval of text chunks, the model gets five tools and digs through the documents itself until it finds. Per the company, accuracy on filings to the US market regulator jumps from 26.7 to 86 percent, without a single line of fine-tuning.
- The model gets five tools over your index: search, open, navigate, read, find an expression inside the open document.
- Per Mistral, accuracy on FinanceBench rises from 26.7 to 86 percent with Mistral Medium 3.5, and token spend drops by 23.9 percent.
- It wants no fine-tuning and works with somebody else's model too. Available through the Mistral Search Toolkit, standalone or in the cloud. No price announced.
I asked a system of ours something that sat in a table deep inside a filing. It gave me back three excerpts of plain text. The table was not among them. The answer was only in it.
The wrong answer came confidently.
That is how ordinary search works under AI. You cut the documents into chunks, the machine finds the closest ones by meaning, hands them to the model and it answers only from them. One move. Miss the chunk on the first try and that is it, the model has no way to ask for another page.
Mistral said it would do it differently: give the model tools and let it dig on its own.
What matters here
Not the number. The number is theirs and was measured by them on their own tool, so I read it as an advert with the workings attached.
The second model in their tests is not theirs at all. GLM-5.2 is Chinese and wins the same. So the mechanism is not tied to a model and wants no fine-tuning. You put it over the index you already have.
That is the quiet part of the whole announcement. For years the quality of a search like this came down to the craft of the person who set up how the PDFs were chunked: how to slice the documents and how much neighbouring text to leave around each piece. You change the model and the system answers almost the same, because the chunks are the same.
With tools it is the other way round. A better model comes along and it simply digs more cleverly through the same documents, without you touching anything.
What is not in the headline
Faster sounds good until you look at what the fast part is. One hundred and fifty-four seconds for a tenth of the questions. Two and a half minutes in which the person watches a spinning circle and does not know whether it is working.
And one more thing. Each of the five tools is a new way for the model to reach a document you did not think it would reach. The index is no longer a display case of selected chunks, it is an entrance to the archive. The rights over that archive become the real setting, not the token count.
There is no price announced. While there is none, the numbers stay a promise.
If you have an archive of documents
Before you change the model over a bad answer, check one thing: did the question even reach the right page. In my case it had not, and no cleverer model was going to save it.
Open one question your system gets wrong and look at what exactly was handed to it. It will take you ten minutes and it will tell you whether you are even looking at the problem in the right place.