Mistral AI · 2026-05-28 · notable
Mistral Search Toolkit — Open-Source RAG Framework With Mistral OCR Ingestion, Hybrid BM25 + Vector Retrieval, Vespa Backing, and a Copier Starter App
A production-oriented RAG framework that bundles document ingestion, hybrid retrieval, semantic caching, and evaluation. Public preview today on PyPI with a copier-scaffolded starter app for Vespa-backed pipelines.
Mistral packages an end-to-end RAG stack — ingestion, retrieval, evaluation — as one Python toolkit with a starter app template.
Key specs
| License | MIT (starter app) |
|---|---|
| Language | Python 3.12+ |
| Backends | Vespa, hybrid BM25 + vector |
| Install | uv add mistralai-search-toolkit |
What is it?
A public-preview framework for building production RAG and search pipelines on top of Mistral's models. It covers configurable ingestion (parse / chunk / embed), hybrid retrieval (BM25 + dense + reranking), evaluation (recall, precision, MRR, NDCG), and a shared interface so components are swappable.
How does it work?
Two orchestration classes do the lifting: `Pipeline` for ingestion and `QueryEngine` for retrieval. Mistral OCR handles PDFs and office files, optional extras add Vespa as the vector store, advanced PDF parsing, spreadsheet support, and cloud storage. A Copier template at `mistralai/search-starter-app` scaffolds a working project with one command.
Why does it matter?
RAG plumbing is the part of every agent stack engineers keep rewriting. Mistral is consolidating ingest-plus-retrieve-plus-eval into one Apache/MIT-licensed package and putting Vespa on the happy path, which is unusual — most open RAG kits default to Chroma or Pinecone.
Who is it for?
Teams building enterprise search, internal knowledge bases, or domain-specific RAG (legal, medical, financial, code).
Try it
uvx copier copy gh:mistralai/search-starter-app my-search-app