Mistral Gives Enterprise AI Five Ways to Keep Digging Through Documents

Mistral has launched Agentic Search for the point where an enterprise AI system finds a plausible document but not the answer inside it. Rather than answer from one set of text excerpts, a model can continue searching, open files, and inspect specific locations before responding.

The launch targets a familiar weakness in retrieval-augmented generation, or RAG. A conventional setup searches an index once, gives a language model a fixed group of text chunks, and asks for an answer. That can work for a direct lookup, but it can break down when the needed figure is in a table, footnote, or a document referenced by the first result. Agentic Search makes the next lookup an action the model can choose while forming its response.

From first result to document trail

The product gives models five tools: search, open, navigate, read, and grep. Together, they let an agent refine a query, open a promising document, move to a page or section, read the material, and look within the file for a term. Mistral says the tools do not require model-specific fine-tuning.

The actions available to the model

• Search: query the indexed document collection for likely sources.
• Open and navigate: inspect a selected document and move to a particular location within it.
• Read and grep: consume relevant content or search inside the file for matching text.

Mistral says the tooling can run in the cloud or on a customer’s premises and connect to an existing search index. Agentic Search is available through Mistral Search Toolkit and through Libraries in Mistral’s Studio and Vibe products. The underlying toolkit entered public preview on May 28 and supports PDFs, office files, spreadsheets, emails, and plain text.

FinanceBench result reported by Mistral

GLM-5.2 accuracy

26.7%→86%

accuracy

Mistral reported this change for GLM-5.2 on FinanceBench.

The early benchmark case

Mistral tested the approach with its Mistral Medium 3.5 model and GLM-5.2, using the Search Toolkit’s default chunking and ranking rather than tuning the system for each use case. On FinanceBench, Mistral says GLM-5.2 rose from 26.7% accuracy with one-shot RAG to 86% with iterative search and document navigation.

FinanceBench evaluates financial question answering across 368 SEC filings covering roughly 53,900 pages. Mistral used its 150-question public evaluation set and scored responses with an LLM judge calibrated against human labels. The company also reported that Mistral Medium 3.5 gained 47.3 percentage points from iterative search and another 8.7 points from navigation.

Mistral says navigation reduced token use compared with a search-only agentic loop by 23.9% for Mistral Medium 3.5 and 33.7% for GLM-5.2. It also reported lower FinanceBench latency: p90 fell from 255 seconds to 154 seconds, while mean latency fell from 108 seconds to 71 seconds.

A second test used OfficeQA Pro, which contains 133 questions over 696 historical U.S. Treasury Bulletins spanning roughly 89,000 pages. Its questions require retrieval and reasoning across scanned tables and multiple documents. Mistral says GLM-5.2 improved from 6.3% with one-shot retrieval to 51.9% with the full toolset.

The next move is a private-corpus test

The results are product evidence, not settled performance. They are Mistral’s own evaluations and have not been independently replicated. The company tested one Mistral model and one third-party model, which offers some cross-checking but does not establish broad model-agnostic performance.

The unresolved question is whether the gains hold in enterprise collections with inconsistent formatting, permissions, and records that do not resemble benchmark material. The launch gives Mistral a product for the retrieval layer; customer deployments will show whether its document-investigation loop stays reliable when finding the right page is only the first obstacle.