Mistral AI Search Toolkit Public Preview

Mistral AI has released Search Toolkit in public preview. This open-source composable framework allows developers to build production search pipelines by unifying ingestion, retrieval, and evaluation under a shared interface, reducing the engineering overhead associated with stitching together disparate tools.

Unified Search Infrastructure

Search Toolkit addresses the fragmentation in current retrieval systems where ingestion, retrieval, and evaluation typically rely on separate toolchains with differing data assumptions. This fragmentation often results in weeks of integration work and makes it is difficult to measure whether a retriever is returning correct results.

Search Toolkit provides a single framework that enables teams to:

  • Standardize Ingestion: Implement consistent processing and indexing patterns across various source types (such as internal wikis, support tickets, and codebases) without rebuilding pipelines for every new source.
  • Isolate Retrieval Quality: Use built-in evaluation tools to measure retriever performance independently from generation quality, allowing teams to determine if a RAG system's failure is due to retrieval or generation.
  • Support Domain-Specific Needs: Provide a structured alternative to building custom retrieval infrastructure from scratch for specialized fields like legal, medical, or financial data.

Technical Components and Capabilities

Search Toolkit is designed to be infrastructure-agnostic, running on cloud, on-premises, or edge environments. It consists of three primary modules that share a common configuration interface:

Ingestion

Search Toolkit manages document parsing, chunking, and embedding generation. It utilizes a standard adapter interface to allow for the integration of custom document formats and preprocessing steps.

Retrieval

The framework supports multiple retrieval configurations, including:

  • BM25 sparse retrieval
  • Dense embedding-based retrieval
  • Hybrid configurations (combining both sparse and dense retrieval)

Evaluation

To track quality across releases and compare configurations, the toolkit includes built-in metrics for measuring search quality, specifically recall, precision, MRR (Mean Reciprocal Rank), and NDCG (Normalized Discounted Cumulative Gain).

Integration with AI Agents

Search Toolkit provides a high-quality indexed search path for AI agents. While agents can use Connectors and MCP integrations to pull live data from CRMs or productivity tools, Search Toolkit allows them to perform low-latency semantic search across large document corpora.

Real-World Application and Implementation

Search Toolkit has been tested across manufacturing, public sector, financial services, and media and entertainment verticals. For example, CMA CGM uses the toolkit alongside Voxtral to detect fake news by processing audio from three different data sources and returning alerts within 15 seconds.

Getting Started

Developers can implement Search Toolkit using a starter app template available on GitHub. The project requires Docker and uv for installation. The template provides:

  • Pre-configured Vespa indexing
  • Hybrid retrieval (BM25 + vector)
  • Sample data and ingestion pipelines

Advanced Optimization

Beyond the basic setup, the framework allows for the following optimizations:

  • Ingestion Tuning: Configuring parsers, chunking strategies, and embedding models for specific file types.
  • Vespa Management: Optimizing indexing and ranking profiles.
  • Retrieval Enhancement: Implementing LLM query rewriting, reranking, and hybrid retrieval.

Sources

Related

  • Dispatch
    Mistral Agentic Search Release

    Mistral AI has introduced Agentic Search, a multi-step retrieval loop that enables AI models to navigate complex, dense documents and improve accuracy on financial and government data benchmarks.

  • Dispatch
    Mistral AI Workflows Public Preview

    Mistral AI has released Workflows in public preview, providing an orchestration layer for enterprise AI that ensures durability, observability, and fault tolerance for production AI processes.

  • Dispatch
    Codestral Embed model release

    Mistral AI released Codestral Embed, a code‑specialized embedding model that outperforms existing code embedders on retrieval benchmarks and supports flexible dimensions and precisions.

  • Dispatch
    Mistral AI Voxtral Release

    Mistral AI has released Voxtral, a family of open-weight speech understanding models in 24B and 3B sizes that provide state-of-the-art transcription and native semantic understanding under the Apache 2.0 license.

  • Project
    vespa-engine/sample-apps

    A collection of reference applications and tutorials for Vespa, demonstrating how to implement hybrid search, RAG, and multimodal vector retrieval at scale.