Mistral has released Agentic Search, turning document retrieval from a one-shot lookup into a multi-step process. The system gives the model five tools: search, open, navigate, read, and grep. The model can run a first search, open a file that matches, jump to a specific section, read it, and if the result looks wrong, revise the keywords and search again — only answering once it's confirmed.
The gains on two benchmarks are substantial. FinanceBench, which draws on SEC filings from U.S. public companies, saw accuracy climb from 26.7% to 86% — more than tripling. OfficeQA Pro, built on U.S. Treasury bulletins, went from 6.3% to 51.9%, a gain of 45.6 percentage points.
Where single-shot retrieval breaks down
Standard RAG works by chunking a document, computing embeddings, retrieving the chunks most similar to the question, and stitching them into the context for the model to answer from. That pipeline is fine for a question like "what was the company's revenue in a given year," but it starts to fail on something like "list free cash flow for the past three years and explain any change in accounting basis."
The failure is baked in at the chunking stage. A single number in a financial statement often needs three things to be read correctly: the value in the table, the footnote beneath it, and an explanation of accounting basis stated in an earlier section. Those three pieces can sit dozens of pages apart, and vector retrieval ranks by semantic similarity — footnotes and basis explanations usually don't rank high enough to make the cut. The model ends up with an isolated number, recites it as-is, and has no way to know if it got it wrong.
Multi-step retrieval restores the process to something closer to how a person flips through a document: locate first, then expand, and when the text says "see Note 12," jump there and come back. The inclusion of grep as a tool is telling in itself: vector search is good at fuzzy semantic matching, but for an exact string like "Note 12," old-fashioned full-text matching is actually more reliable. The two retrieval methods each handle their own part within the same toolset.
There's an easy-to-miss benefit buried here too. A single-shot answer can't be traced back — if the model says revenue was a certain figure and the user wants to know which page it came from, they have to go dig through the document themselves. Every step of multi-step retrieval is an explicit tool call, and which file was opened, which section it jumped to, which string it matched — all of that stays in the trace. For finance and legal use cases, that trail is itself part of the deliverable: getting the answer right isn't enough, you also need to be able to show why it's right.
More passes, yet cheaper
Intuitively, multi-step operation should mean more rounds, more tokens, and longer waits. Mistral's numbers go the other way: token usage drops by up to 33.7%, and P90 latency falls 39.6%, taking FinanceBench from 255 seconds down to 154 seconds.
These numbers hold only relative to a specific baseline. To guard against missing anything, single-shot RAG tends to widen the retrieval count — stuffing twenty or eighty chunks into the context, most of which have nothing to do with the question. Multi-step retrieval reads only the segment it needs on each pass; there are more passes, but each one's input is short, and the total ends up smaller.
The latency figure depends more on the engineering. A lower P90 means fewer of the tail cases where "retrieval fails, the model fabricates an answer, the result is way off, and it has to be rerun." A single-shot failure is silent — the model has no idea it never got the footnote. In multi-step retrieval, the model can notice it hasn't found what it needs and search again, which narrows the tail instead of widening it.
Two hurdles for Chinese-language use
Mistral ships this as a Search Toolkit and Libraries, already integrated into Studio and Vibe, and it supports both cloud and on-premises deployment. On-prem is a hard requirement for industries like finance, healthcare, and government that can't let documents leave their premises, and Mistral has consistently positioned it as a point of differentiation from U.S. vendors.
Porting this to Chinese-language documents runs into two new problems. First, Chinese text has no word boundaries, so grep's exact-match advantage shrinks compared with English, and a rough estimate suggests match noise would run noticeably higher. Second, domestic prospectuses, annual reports, and policy documents rely heavily on scanned images and non-standard tables — whether the read step can reconstruct table structure directly determines whether the later steps are reading or guessing.
Another caveat concerns the benchmarks themselves. FinanceBench and OfficeQA Pro are both public question sets — their questions and source documents have long circulated online, so any vendor running its own system against them risks training-data contamination. The 26.7% starting point is strikingly low, and the announcement doesn't break down which RAG configuration it corresponds to, how many chunks were retrieved, or which base model was used. The direction of the improvement is credible; the magnitude should be taken with a discount.
Pricing hasn't been announced yet. That matters a lot for enterprise buyers: each multi-step query triggers multiple model calls, and if billing is per call, getting to that 86% accuracy could cost several times the per-query price — the math has to be redone.
Sources: Mistral's official announcement, CocoLoop; the FinanceBench and OfficeQA Pro accuracy figures, token savings, and P90 latency numbers were checked line by line.