Mistral AI has introduced Search Toolkit in public preview, positioning it as an open-source framework for building production search systems for AI applications.
The company said the toolkit is designed to reduce the engineering work involved in search infrastructure, where teams often rely on separate tools for ingesting data, retrieving results and evaluating quality. By combining those functions in a single framework, Mistral aims to let developers spend less time on integrations and more time improving search performance.
Search Toolkit is built for deployment across different environments, including cloud systems, on-premises infrastructure and edge setups. Mistral said the product is intended to be composable, so teams can swap components without rebuilding an entire pipeline.
According to Mistral, many teams building retrieval systems spend significant time assembling the underlying plumbing before they can begin testing search quality on their own data. The company said ingestion, retrieval and evaluation often come from different frameworks with different assumptions about data structure, which can slow down development and make experimentation harder.
Mistral is pitching Search Toolkit as a way to unify those stages under one interface. The company said this should help organisations work with internal knowledge systems, retrieval-augmented generation workflows and other search-heavy applications without stitching together multiple disconnected tools.
The toolkit supports enterprise search use cases where organisations may need to index content from many sources, such as internal wikis, support systems, file repositories and codebases. Mistral said the framework provides consistent patterns for processing and indexing different source types, which could reduce the need to build a custom pipeline each time a new source is added.
Search Toolkit includes several retrieval options, including BM25 sparse search, dense embedding-based retrieval and hybrid configurations that combine both approaches. Mistral said the modules share a common configuration interface, allowing teams to adjust or replace parts of the pipeline without rewriting the rest.
The toolkit also includes built-in evaluation tools. Mistral said these can measure metrics such as recall, precision, mean reciprocal rank and NDCG against a team’s own test sets. That is meant to help developers compare retrieval strategies directly and separate retrieval quality from generation quality when diagnosing problems in RAG systems.
The company also highlighted the toolkit’s use in domain-specific environments such as legal, medical, financial and code retrieval, where general-purpose search systems may struggle with specialised terminology or document structure.
Mistral said Search Toolkit is relevant for agentic systems that need enterprise context to complete tasks. In those workflows, agents may need to search a large indexed corpus for low-latency semantic results, while also pulling live data from source systems when up-to-date information is required.
The company pointed to its Connectors feature, which lets agents retrieve live information from systems such as CRMs, code repositories and productivity tools through MCP integrations. In Mistral’s framing, the toolkit is meant to give agents both indexed search and live retrieval paths.
To help developers begin testing the product, Mistral is offering a starter app template with preconfigured Vespa indexing, hybrid retrieval and sample ingestion pipelines. The template can be used with Docker and uv, and the company provided commands for setting up a local Vespa instance, ingesting sample data and running a query.
Mistral said Search Toolkit has already been tested in sectors including financial services, manufacturing, the public sector and media and entertainment. It gave one example of CMA CGM using the toolkit together with Voxtral to help journalists identify fake news, with alerts reportedly returned in 15 seconds end to end.
The public preview adds another tool to Mistral’s broader enterprise product lineup as AI developers continue to look for simpler ways to build and evaluate search systems that underpin RAG applications and autonomous agents.