Introducing Search Toolkit

Mistral points out that most teams spend more time integrating infrastructure than improving search quality when building retrieval systems. Data ingestion, retrieval, and evaluation each use different tools and interfaces, causing teams to take weeks to run their first query on their own data. For organizations building RAG workflows or internal knowledge systems, this overhead compounds.

Search Toolkit is a composable framework that unifies data ingestion, retrieval, and evaluation under a shared interface. It addresses the common 'multiple search problems' in enterprise search—internal wikis, ticketing systems, document repositories, etc., each with different structures and metadata, often requiring separate ingestion pipelines per source. Search Toolkit provides consistent processing and indexing patterns across source types, enabling teams to add new sources without rebuilding pipelines.

In RAG scenarios, when the system returns poor results, teams often cannot tell whether the issue is retrieval or generation. Search Toolkit's built-in evaluation can independently measure retriever performance, thus isolating retrieval quality from generation quality, and supports comparing different configurations on your own data.

For specialized domains like legal documents, medical records, or codebases, generic retrievers often perform poorly. Search Toolkit allows teams to build domain-customized retrieval infrastructure without starting from scratch.

Mistral emphasizes that agents working on enterprise tasks need access to enterprise context; they make retrieval decisions autonomously and at high frequency, so the quality of the underlying search infrastructure directly impacts every downstream step. Search Toolkit provides high-quality indexed search paths for agents, while supporting real-time data pulling from source systems like CRM and code repositories via Connectors and MCP integration.

Search Toolkit has been battle-tested in finance, manufacturing, public sector, and media/entertainment industries. For example, CMA CGM combined Search Toolkit with Voxtral to help journalists detect fake news: the pipeline processes audio from three different data sources and returns alerts within 15 seconds.

Users can get started quickly with the starter app template, provided Docker and uv are installed.