Search Toolkit
Search Toolkit is a Python framework for building information retrieval (IR) systems. It provides components for ingestion, retrieval, and evaluation. It works across backends, and every component can be swapped.
LLMs aren't trained on your private data. To ground their answers in your documents, you need a retrieval pipeline that ingests, indexes, and searches your content. Search Toolkit provides the building blocks for that pipeline.
Key features
Ingestion
Ingestion includes:
- Multi-format extraction: PDF, DOCX, and PPTX through Mistral OCR, HTML, spreadsheets, emails, plain text
- File loading: Load from the local filesystem, S3, Google Cloud Storage, or Azure Blob Storage, or implement custom loaders for any source
- Flexible chunking: Character, token, markdown-aware, or separator-based splitting
- Enrichment: Enrich documents and chunks with custom metadata or LLM-generated summaries
- Indexing: Index to vector stores for semantic search
Retrieval
Retrieval includes:
- Multiple strategies: Vector, keyword, and hybrid search, fused with RRF, with optional reranking
- Query preprocessing: Improve user queries with LLM reformulation, or generate sub-queries with query extension
- Reranking: LLM reranker, cross-encoder reranker, or custom rerankers
- Semantic caching: Cache results by query similarity to skip redundant retrieval
Architecture
Ingestion transforms raw documents into searchable chunks. A FileLoader reads raw bytes from a source, a DocumentExtractor converts them into a structured document, a TextSplitter divides it into chunks, optional DocumentProcessor stages (such as a ChunkEnricher or SummaryEnricher) add metadata, and an Embedder produces vectors for indexing into a vector store.
Retrieval finds relevant chunks for a given query. An optional query rewriter rewrites the query, a Retriever searches the index, and an optional Reranker re-scores the results before returning them. An optional query extension generates sub-queries that are returned with the result but not searched.
Both workflows are orchestrated by high-level classes (Pipeline for ingestion, QueryEngine for retrieval) that handle component wiring and execution. Every component can be swapped: use the built-in implementations or bring your own.
Components
| Component | Built-in options |
|---|---|
| File loaders | FilesystemFileLoader, S3FileLoader, GCSFileLoader, AzureBlobFileLoader, custom loaders |
| Extractors | MistralOCRExtractor, SmartPDFExtractor, PlainTextExtractor, PyMuPDFExtractor, HTMLExtractor, SpreadsheetExtractor, EmailExtractor, EmailWithAttachmentsExtractor, NumbersExtractor, LegacyOfficeExtractor, MistralAudioTranscriptionExtractor |
| Text splitters | CharacterTextSplitter, TokenTextSplitter, MarkdownTextSplitter, ParagraphTokenTextSplitter, MarkdownTokenTextSplitter, SeparatorTextSplitter |
| Enrichers | SummaryEnricher, custom ChunkEnricher |
| Embedders | MistralEmbedder, custom Embedder |
| Storage | Mistral Search Index, Vespa, Postgres (pgvector), Qdrant, or custom vector store |
| Retrievers | VectorRetriever, KeywordRetriever with a custom KeywordStoreIndex |
| Rerankers | LLMReRanker, CrossEncoderReRanker, RRFRanker |
| Preprocessing | LLMQueryRewriter, LLMQueryExtension |
| Caching | SemanticCache with InMemoryCacheBackend |
Installation and extras
Install the core package:
uv add mistralai-search-toolkitOptional extras add specialized functionality:
| Extra | Description |
|---|---|
vespa | Vespa plugin for vector storage and semantic search |
postgres | PostgreSQL backend with pgvector and pg_textsearch |
extractor-pymupdf | Advanced PDF extraction with PyMuPDF Pro |
extractor-spreadsheet | Spreadsheet parsing (Excel, CSV, Apple Numbers) |
extractor-email | Email file parsing (EML, MSG formats) |
html-converter-markdownify | Convert HTML to Markdown |
text-splitter-langchain | Additional text splitting strategies via LangChain |
storage-gcs | Google Cloud Storage integration |
storage-azure | Azure Blob Storage integration |
storage-s3 | Amazon S3 integration |
Install extras with the core package:
uv add "mistralai-search-toolkit[vespa]"
uv add "mistralai-search-toolkit[vespa,extractor-pymupdf]"All packages are available on PyPI.
Requires Python 3.12 to 3.14. We recommend uv for dependency management.
Next steps
- Quickstart: build your first ingestion and retrieval pipeline end to end.
- Document model: understand
Document,DocumentChunk, and the identity that ties them together. - Embedding model: understand how the schema describes embedding vectors, separately from the embedder.
- Search index: set up your vector store.
- Ingestion: load, extract, chunk, enrich, and index your documents.
- Retrieval: configure vector search with optional reranking.