NextLUCA
← ← Back to blog

Mistral Agentic Search: A Multi-Step Retrieval Loop for Document-Heavy Work

Mistral AI has detailed Agentic Search, a retrieval system built around a multi-step loop of finding, inspecting, and verifying information across data sources. It's available through the Mistral Search Toolkit and built into Libraries in Studio and Vibe, and the benchmark numbers behind it are specific enough to be worth walking through.

Mistral AI has detailed Agentic Search, a retrieval system built around a multi-step loop of finding, inspecting, and verifying information across data sources. It's available through the Mistral Search Toolkit and built into Libraries in Studio and Vibe, and the benchmark numbers behind it are specific enough to be worth walking through.

  • A multi-step retrieval loop that can find, inspect, and verify information across data sources, rather than a single retrieval pass.
  • Five distinct tools exposed to the agent: search, open, navigate, read, and grep.
  • Available through the Mistral Search Toolkit, and built into Libraries within Studio and Vibe.
  • On FinanceBench (368 SEC filings, 150 questions), correctness rose from 26.7% to 86%.
  • On OfficeQA Pro (696 Treasury Bulletins, 133 questions), accuracy rose from 6.3% to 51.9%.
  • Adding the navigate tool alone cut FinanceBench p90 latency from 255 to 154 seconds, and mean latency from 108 to 71 seconds.
  • Across benchmarks, p90 latency dropped by up to 39.6% and token consumption dropped by up to one-third.
  • Teams building retrieval workflows over large sets of structured filings or bulletins, similar in scale to the SEC filings and Treasury Bulletins used in these benchmarks.
  • Developers already using Studio or Vibe who want retrieval tools with more steps than a single search-and-answer pass.
  • Anyone evaluating retrieval accuracy and latency trade-offs on document-heavy question-answering tasks.

Access Agentic Search through the Mistral Search Toolkit, or use it as it's built into Libraries in Studio and Vibe. To evaluate it on your own data, replicate the benchmark setup Mistral published: assemble a document set (their tests used SEC filings and Treasury Bulletins), write a fixed set of test questions, and compare correctness, p90 latency, and token consumption with and without the navigate/grep tools enabled, the same variables Mistral reports changing FinanceBench results from 26.7% to 86% correctness.

Sources