product update

Mistral Launches Agentic Search, Claims 3x Accuracy Gain on Financial Document Retrieval

TL;DR

Mistral has released Agentic Search, a retrieval layer that lets AI models navigate, read, and verify information across complex documents instead of relying on single-pass chunk retrieval. The company claims accuracy improvements from 26.7% to 86% on FinanceBench and up to 39.6% lower p90 latency.

3 min read
0

Mistral has launched Agentic Search, a retrieval layer designed to let AI systems navigate, read, and verify information inside complex documents rather than answering from a fixed set of retrieved text chunks. The tool is available now through Mistral Search Toolkit and is built into Libraries in both Studio and Vibe products.

According to Mistral, Agentic Search replaces traditional one-shot RAG (retrieval-augmented generation) with a multi-step retrieval loop. Instead of a model answering from an initial set of top-k chunks, it can search, inspect, refine, open documents, navigate to specific sections, and read source material before responding.

Benchmark claims

Mistral reports the following results, though these figures are self-published and not independently verified:

  • FinanceBench: accuracy improved from 26.7% to 86% — roughly a 3x gain in correctness on financial filing questions.
  • OfficeQA Pro: a +45.6 point gain, from 6.3% to 51.9%, on table-heavy, multi-document questions.
  • Latency: p90 latency reduced by up to 39.6% through targeted navigation.
  • Token usage: reduced by up to one-third due to fewer repeated searches.

No pricing details were disclosed for Agentic Search as a standalone offering; it is packaged within existing Mistral Search Toolkit and Libraries products.

How it works

Agentic Search builds on top of an existing search index and adds five tools that resemble file-system operations:

  • search — finds relevant documents across a corpus
  • open — opens a specific document
  • navigate — moves to a page, section, or region within a document
  • read — retrieves content at a given location
  • grep — finds a pattern within an open document

Mistral says these tools require no fine-tuning or model-specific training, meaning retrieval quality should improve as underlying models get better at reasoning and tool use, rather than being capped by a fixed chunking strategy.

The company illustrates the difference with a worked example: given a question about total U.S. national defense expenditures across all months of 1953, one-shot retrieval using only a search call returned partial-year data (January–June) scattered across fiscal-year-framed bulletins, an incomplete answer. Agentic Search, using two search calls followed by a read call, located a complete table covering all 12 months and computed a sum of $44,463 million.

Where it applies

Mistral positions Agentic Search for long documents (filings, contracts, technical specifications), multi-source questions requiring comparison or reconciliation across documents, answers that require verifiable citation to a specific document location, and tables or structured/scanned documents where meaning depends on layout rather than narrative text. The company recommends standard indexed retrieval for simpler cases: direct lookups in short documents, high-volume keyword/semantic search, and predictable questions where the answer location is already known.

Mistral also emphasizes that the toolkit is designed to operate within isolation boundaries — cloud or on-premises — for enterprises handling sensitive domain-specific data such as financial filings, legal contracts, and government records.

What this means

Agentic Search is Mistral's answer to a well-known limitation of RAG systems: single-pass retrieval fails on dense, multi-document, or table-heavy content where the answer isn't cleanly contained in a top-ranked chunk. The reported jump from 26.7% to 86% on FinanceBench is a self-reported figure from Mistral and hasn't been independently benchmarked, so enterprises evaluating this should test it against their own document sets rather than take the headline number at face value. The more durable claim is architectural: by decoupling retrieval quality from chunking strategy and tying it instead to model reasoning and tool-use capability, Mistral is betting that this approach improves automatically as its underlying models get better — without requiring re-engineering of the retrieval pipeline. That's a meaningful shift for enterprises with large, sensitive document repositories in finance, legal, and government sectors, where existing RAG setups routinely fail on tables, footnotes, and cross-document references.

Related Articles

product update

Mistral launches Vibe Desktop for macOS and Windows, running parallel agent tasks locally or in the cloud

Mistral has released Vibe Desktop, a desktop app for its Vibe agent (formerly Le Chat), for macOS and Windows. It runs multiple tasks in parallel in a local repository or a cloud sandbox, and a Linux version is planned.

product update

Microsoft makes Word, Excel and PowerPoint in Copilot generally available, calls Copilot "the new OS for work"

Microsoft CEO Satya Nadella announced Tuesday at Ignite the general availability of Word, Excel and PowerPoint inside Copilot, which he called "the new OS for work." The company also introduced Copilot Autopilot, Microsoft IQ and expanded Agent 365 controls, with pricing for Autopilot still unannounced.

product update

Replit launches Missions: agent swarms work toward goals within budgets and approval limits you set

Replit has launched Missions, a product where users define an outcome and Replit coordinates a swarm of agents to plan, build, test and iterate within a budget and boundaries the user sets. Replit says it used Missions to build an application monitoring platform, DataCat, in 2 days within a $3,000 budget.

product update

Perplexity Brain Can Now Create and Update Agent Skills; 80,000+ Personal Skills Written for 20,000+ Users

Perplexity said on November 18, 2026, that Brain, the personalization system for its Computer agent, can now create and update reusable agent skills. According to the company, Brain wrote more than 80,000 personal skills for over 20,000 users between August 1, 2026, and November 14, 2026.

Comments

Loading...