Mistral Introduces Agentic Search, Tripling Accuracy on Financial Filings

Mistral launched Agentic Search, a retrieval capability designed to make AI systems dramatically more accurate and efficient when working through complex documents. Available via Mistral's Search Toolkit and Libraries, it lets an AI navigate, read, and verify documents with targeted navigation rather than brute-force repeated searches.

The reported numbers are the story. On FinanceBench — a benchmark of questions against real financial filings — Agentic Search lifts accuracy from 26.7% to 86%, roughly a 3x improvement, a domain notorious for dense, cross-referenced tables and footnotes. It also cuts token consumption by up to one-third and reduces p90 latency by up to 39.6%, meaning the slowest tail of queries gets meaningfully faster. The mechanism is agentic: rather than dumping retrieved chunks into context, the system reasons about where to look and verifies as it goes, reducing both wasted tokens and repeated round-trips.

This lands in a busy week for retrieval and document-AI economics. AWS separately described query-aware context compression on Bedrock to cut RAG costs, and LlamaIndex CEO Jerry Liu published findings that coding-agent harnesses like Claude Code are surprisingly strong cost/accuracy baselines for long-document extraction. The convergence signals that retrieval quality — not just model size — is the current competitive battleground for enterprise document workflows.

The caveat: vendor-reported benchmarks warrant independent replication, and FinanceBench gains may not generalize to messier real-world document sets. Still, for regulated industries drowning in filings, a verified 3x accuracy jump at lower cost is a compelling pitch. Watch for third-party FinanceBench reproductions and whether Agentic Search integrates with Mistral's sovereign-cybersecurity work for France.