Mistral AI launches Agentic Search for complex document retrieval

Mistral AI launched Agentic Search on August 20, a multi-step retrieval layer that lets AI systems search, inspect and verify information across complex documents before producing an answer. It is available through Search Toolkit, MCP tools and Libraries in Studio and Vibe.

One-shot retrieval-augmented generation hands a model an initial set of text chunks and asks it to answer from those. Agentic Search can repeat the retrieval process instead, refining a query, inspecting multiple sources and excluding chunks it has already seen while it gathers evidence.

The workflow runs on an existing semantic or hybrid index and gives the model five document operations: search, open, navigate, read and grep. According to Mistral’s product documentation, those operations let an agent find candidate documents, expand context around a result, move through a document, retrieve content from a known range and locate exact terms.

Developers can wire the workflow into custom agents with the Search Toolkit, a Python framework covering ingestion, retrieval and evaluation, or expose its operations to MCP-compatible agents. Mistral’s Libraries hold persistent document collections that agents can query through a built-in document tool with source references. The launch follows another recent addition to Mistral’s product lineup, the Shieldstral moderation model.

Mistral also reported benchmark gains, though those results have not been independently reproduced in the supplied evidence. On FinanceBench, the company said correctness rose from 26.7% with one-shot retrieval to 86% with the full workflow, p90 latency fell from 255 seconds to 154 seconds, and navigation reduced token use by as much as 33.7%. Mistral said its tests used the default Search Toolkit stack without tuning and evaluated Mistral Medium 3.5 and Z.ai GLM-5.2.