Mistral launches Agentic Search so enterprise AI can keep looking

Mistral AI launched Agentic Search on August 20, giving enterprise AI systems a multi-step retrieval loop to search, open, and read long documents instead of answering from the first batch of text chunks. In a company-run test on FinanceBench, Mistral reported that GLM-5.2 accuracy rose from 26.7% with one-shot RAG to 86% with iterative search and document navigation. The product is available through Mistral Search Toolkit and Libraries inside Studio and Vibe, and can run in the cloud or on-premises.

Mistral launches Agentic Search so enterprise AI can keep looking Arthur Mensch's Mistral says its retrieval loop lifted FinanceBench accuracy from 26.7% to 86% in a company-run test. By RuntimeWire Staff /author/runtimewire-staff · Published Primary source: Mistral AI https://mistral.ai/news/agentic-search/ Why it matters Enterprise agents fail when they cannot locate and verify private information. Mistral is turning retrieval into a core product as it builds beyond foundation models. Arthur Mensch's Mistral AI https://mistral.ai/?ref=runtimewire launched Agentic Search https://mistral.ai/news/agentic-search/?ref=runtimewire on August 20th, giving enterprise AI systems a multi-step retrieval loop for searching, opening and reading long documents instead of answering from the first batch of text chunks they receive. Mensch, a former Google DeepMind researcher, founded Mistral in 2023 with Guillaume Lample and Timothee Lacroix, who had worked at Meta AI. The three researchers built Mistral around a pitch of efficient models, open tooling and customer control. Agentic Search carries that argument into the retrieval layer, where access to private information often matters as much as the model generating the final answer. The product is available through Mistral Search Toolkit https://docs.mistral.ai/studio/search-toolkit?ref=runtimewire and through Libraries https://docs.mistral.ai/studio/libraries?ref=runtimewire inside Studio https://mistral.ai/products/studio/?ref=runtimewire and Vibe https://mistral.ai/products/vibe/?ref=runtimewire . Mistral says customers can run the tooling in the cloud or on-premises and connect it to an existing search index, an important requirement for organizations that cannot ship contracts, filings or operational records to a third-party service. A model that can keep looking Conventional retrieval-augmented generation, or RAG, usually searches an index once, passes a fixed group of chunks to a language model and asks for an answer. That works for a direct lookup. It becomes brittle when the relevant figure sits in a footnote, a table on page 147 or a second document referenced by the first. Agentic Search lets the model decide what to inspect next through five tools: search , open , navigate , read and grep . The model can refine a query, open a promising document, move to a particular page and search within that file before responding. The tools require no model-specific fine-tuning, according to Mistral. Mistral illustrates the difference with a question asking for the sum of monthly U.S. national defense expenditures in 1953. A single search retrieved bulletins containing only part of the year. The agentic loop searched again, found a February 1954 Treasury Bulletin containing all 12 months, read the relevant page and calculated a total of $44,463 million. That example captures the practical advantage of the approach. The initial index still narrows the corpus. The model then investigates likely sources rather than treating the first search results as the entire available record. The underlying Search Toolkit https://mistral.ai/news/search-toolkit/?ref=runtimewire entered public preview on May 28th. It handles document extraction, chunking, embeddings, indexing, retrieval and evaluation across formats including PDFs, office files, spreadsheets, emails and plain text. Agentic Search turns that infrastructure into a set of actions a model can call during inference. The benchmark case Mistral tested the system with Mistral Medium 3.5 /models/mistralai/mistral-medium-3-5 and Z.ai's GLM-5.2 /models/z-ai/glm-5.2:free on FinanceBench and OfficeQA Pro. The tests used Search Toolkit's default chunking and ranking without use-case-specific tuning, according to Mistral. On FinanceBench, Mistral reports that GLM-5.2 accuracy rose from 26.7% with one-shot RAG to 86% with iterative search and document navigation. The search loop accounted for a 52.6 percentage-point increase, while the open , navigate , read and grep tools added another 6.7 points. Mistral Medium 3.5 gained 47.3 points from iterative search and another 8.7 points from navigation. FinanceBench https://arxiv.org/abs/2311.11944?ref=runtimewire evaluates financial question answering across 368 SEC filings covering about 53,900 pages. Mistral used its 150-question public evaluation set and scored answers with an LLM judge calibrated against human labels. The navigation tools also reduced the cost of repeated searching in Mistral's test. Token use fell 23.9% for Mistral Medium 3.5 and 33.7% for GLM-5.2 compared with a search-only agentic loop. FinanceBench p90 latency declined from 255 seconds to 154 seconds, while mean latency fell from 108 seconds to 71 seconds. On OfficeQA Pro https://arxiv.org/abs/2603.08655?ref=runtimewire , GLM-5.2 rose from 6.3% under one-shot retrieval to 51.9% with the full toolset, Mistral says. The benchmark contains 133 questions over 696 historical U.S. Treasury Bulletins spanning roughly 89,000 pages. Its questions require systems to retrieve and reason over scanned tables and multiple documents, making correct document selection only the beginning of the task. The gains are Mistral's own evaluation rather than independently replicated results. Testing a Mistral model and a third-party model provides a useful cross-check, although two models do not establish Mistral's broader claim that the system is model-agnostic. The OfficeQA Pro research itself found that frontier agents supplied with the document corpus averaged 34.1%, evidence that retrieval and document parsing remain major failure points even when capable models receive the underlying files. Mensch moves up the enterprise stack Agentic Search places Mistral in a market already occupied by specialists. Glean https://www.glean.com/enterprise-search?ref=runtimewire sells permissions-aware workplace search across enterprise applications, while Contextual AI https://contextual.ai/blog/contextual-ai-platform-generally-available?ref=runtimewire offers active, multi-step retrieval for specialized enterprise agents, including private deployment options. Mistral's pitch rests on portability, open components and compatibility with indexes customers already operate. That fits Mensch's longer campaign to give European organizations greater control over the systems running their data. Mistral's founders describe openness, efficiency and user control as core to the Paris-based AI developer's founding thesis. The launch also extends a rapid move beyond foundation models. RuntimeWire reported earlier in August that Mistral shipped a compact guard model with runtime policy controls /article/mistral-shieldstral-3b-runtime-policy-safety-model , and that Mensch is building a European compute coalition around long-term capacity commitments /article/mistral-european-compute-units-ai-capacity-coalition . Search adds another layer between Mistral's models and the enterprise workflow. That layer can determine whether an agent is useful. A stronger model cannot produce a verified answer when its retrieval system sends it the wrong page. Mensch is betting that Mistral can sell the machinery that finds the evidence, applies the policy and runs inside the customer's infrastructure, rather than leaving those jobs to a collection of outside vendors. Agentic Search gives Mistral a credible entry into that layer. The benchmark results show that letting a model investigate documents can produce a substantial gain over a single retrieval pass. Enterprise deployments will determine whether those gains survive private corpora with inconsistent formatting, access controls and fewer convenient benchmark answers.