Mistral introduces Agentic Search

What changed

Source summary

Mistral launched Agentic Search, a retrieval system giving AI models five tools (search, open, navigate, read, grep) to iteratively navigate, verify, and cross-reference information across complex documents rather than relying on one-shot retrieval. Mistral reports accuracy gains on document-heavy benchmarks (e.g. FinanceBench from 26.7% to 86%), up to 39.6% lower latency, and roughly one-third lower token consumption. It is available through Mistral's Search Toolkit and integrated into Studio and Vibe.

Why it matters to your work

Practical interpretation

If your team uses Mistral's tools to search or summarise dense documents — contracts, financial reports, compliance material — this should produce more accurate, better-verified answers with less waiting and lower running costs, especially for documents too long for a single-pass search to handle well.

When this could help

Worth evaluating if your RAG application handles complex multi-page documents and your users report missed or incomplete answers — especially in finance, legal, or regulatory domains.

Mistral Agentic Search is a multi-step retrieval loop that goes beyond standard RAG by searching, opening, navigating, reading, and verifying information across documents — comparing multiple sources rather than just returning top-k chunks. On FinanceBench (368 SEC filings), correctness rose from 26.7% to 86% (~3x). On OfficeQA Pro (696 Treasury Bulletins), scores improved by up to 45.6 percentage points. P90 latency fell up to 39.6% and token use fell up to one-third.

How to get started

1. For quick testing: use the Search Starter App from GitHub to create a local index. For production: integrate via the Mistral Search Toolkit, configuring ingestion (parsers, chunking, embeddings), indexing (Vespa schemas, relevance profiles), and retrieval extensions (query rewriting, reranking, hybrid retrieval).

   Provider guidance ↗

Build it into your routine

SCALA suggested workflow

1. Complex document QA upgrade

   If your existing RAG pipeline struggles with long documents, test Mistral Agentic Search on the same query set and compare correctness rates to see if the multi-step approach closes the gap.

What to check

• One-shot RAG is often sufficient for direct lookups in short, clean documents. Agentic Search adds value mainly on complex, multi-source queries. Source ↗
• Results depend on the harness: GLM-5.2 scored 41.4% on OfficeQA Pro with the Claude Code harness vs 51.9% with Mistral's. Pricing, rate limits, and data retention are not documented in the announcement. Source ↗