Mistral has replaced single-pass document retrieval with a five-tool loop that lets a model work through a filing the way an analyst would, not the way a search engine returns a snippet. The company calls the product Agentic Search, and it now ships inside Mistral Search Toolkit as well as Libraries in Studio and Vibe.
The mechanism matters more than the branding. A conventional retrieval-augmented generation setup pulls a fixed set of chunks once and asks the model to answer from whatever it got, even when the passage is incomplete or pulled from the wrong section. Agentic Search instead hands the model five operations that resemble file-system commands: search, open, navigate, read, and grep. A model can decide a first search missed the right table, open a different document, jump to a specific page, and grep for a pattern before it commits to an answer.
Mistral’s own benchmarks describe the gap this closes. On FinanceBench, a 150-question test built from 368 SEC filings averaging roughly 147 pages each, correctness rose from 26.7 percent under one-shot retrieval to 86 percent with the full agentic loop. Mistral credits most of that jump to the search step alone: a 47.3 point gain for its own Mistral Medium 3.5 model and a 52.6 point gain for Z.ai’s GLM-5.2, with the open, navigate, read, and grep tools adding another 6.7 to 8.7 points on top. On OfficeQA Pro, a harder benchmark built from 696 scanned Treasury Bulletin PDFs, GLM-5.2 climbed from 6.3 percent to 51.9 percent.
These are vendor-run figures. Mistral built the benchmark harness, chose the default chunking and ranking settings, and published the results on its own blog with no independent replication cited. A jump from 26.7 to 86 percent correctness is the kind of swing that normally draws outside scrutiny before anyone treats it as settled. Mistral does supply one partial check on itself: citing Kimi research, it notes GLM-5.2 scores 41.4 percent on OfficeQA Pro when run through the Claude Code harness, against 51.9 percent on Mistral’s own harness, a 10.5 point gap on the identical underlying model. That gap is itself evidence the numbers move with the measurement tooling, not only with the retrieval method.
Mistral also reports operational gains that matter to anyone running retrieval in production. On FinanceBench, p90 latency fell from 255 seconds to 154 seconds, and the full loop used up to a third fewer tokens than a search-only loop by trading repeated broad queries for targeted navigation.
The architectural argument is the more durable part of this release. Plain chunk retrieval treats the index as the final word: whatever the top results return is what the model gets, and if the real answer sits in a footnote three pages past the retrieved chunk, the system fails quietly. Agentic Search treats the index as a starting point for a loop the model can direct, closer to how a person actually checks a cited figure inside a long report.
Anyone currently running one-shot RAG against long, table-heavy, or multi-document sources (filings, contracts, scanned government records) should read this as a reason to prototype an agentic loop against their own eval set, not to accept Mistral’s numbers as portable. The 10.5 point swing between two harnesses on one model is the clearest signal that results will shift with your retrieval stack, not just your choice of model.
Mistral detailed Agentic Search and its FinanceBench and OfficeQA Pro results in a blog post on its own news site, mistral.ai, this week.
