Mistral announced Agentic Search, a retrieval layer that lets models search, open, navigate, read and grep inside complex documents instead of answering only from the initial chunks of classic RAG. ModelDex and Mistral’s blog highlighted the core claim: roughly triple accuracy on financial documents.
On FinanceBench — 368 SEC filings and 150 questions, documents averaging about 147 pages — Mistral reports Medium 3.5 rising from 26.7% to 86% accuracy with the full agentic loop. That jump of more than 59 percentage points underpins the “triples accuracy” message.
The lab also measured OfficeQA Pro, a set of scanned, table-heavy U.S. Treasury bulletins. There, with GLM-5.2, the full loop reaches 51.9% versus 6.3% for one-shot RAG (+45.6 points). Figures are Mistral-reported; they are not an independent Neural Pulse audit.
Agentic Search builds on an existing index and exposes five tools to the model: search, open, navigate, read and grep. The index flags candidates; the model decides what to inspect, refining searches when the first hit is incomplete — the typical one-shot RAG failure on tables and footnotes.
Beyond accuracy, Mistral claims lower latency and fewer tokens: on FinanceBench, adding navigation cut p90 latency (255s to 154s in its tables) and token use versus loops that only repeat broad searches. Fewer blind retries, more targeted inspection.
The product ships via Mistral Search Toolkit — integrable into customers’ own agents, in cloud or on-premises — and inside Libraries in Studio and Vibe. Emphasis on isolated deployments answers demand from banks and administrations that cannot send sensitive documents to public APIs.
For enterprises, the practical meaning is that financial, legal or operations agents can cite a specific page and table instead of paraphrasing a truncated chunk. That matters when a figure error carries regulatory or audit cost.
Limits remain real: the loop consumes turns and time; a well-built index is required; and vendor benchmarks must be validated on one’s own corpus. Third-party evaluations already warn that reproducing FinanceBench does not guarantee the same delta on messy internal PDFs.
In the 2026 enterprise-agent race, OpenAI, Anthropic and Microsoft push orchestration; Mistral bets on portable, open retrieval tools that scale with model capability without retraining the index. If the model reasons better tomorrow, the same toolkit should yield more.
Neural Pulse highlights the claim as Mistral states it — triples accuracy on financial filings, 26.7% to 86% on FinanceBench with Medium 3.5 — and the product significance for enterprise RAG/agents, without inventing extra metrics.
Data teams should trial Agentic Search on a labeled subset of their own 10-Ks, contracts or manuals before replacing RAG pipelines. Gains usually concentrate on multi-document questions and tables, not one-sentence lookups.
October 2026 finds Mistral selling verifiable precision on dense documents as a European edge against generalist assistants. In finance, “almost right” is not enough; Agentic Search tries to make the agent show where it read the number.
