MiniRAG
Provide local-first hybrid search over personal and project documents with semantic embeddings and BM25 keyword matching.
MiniRAG is a privacy-first retrieval-augmented generation (RAG) MCP server designed to search and index local documents without exposing data to the cloud. Leveraging LanceDB, FastEmbed, and Reciprocal Rank Fusion, it combines deep semantic vector similarity with exact BM25 keyword matching to surface relevant code snippets, error identifiers, and text passages. It supports multi-format parsing across 12 document types via markitdown, optional local OCR for scanned PDFs, and can serve multiple distinct project indexes from a single server or delegate embedding workloads to a local llama.cpp endpoint.
主な機能
01Comprehensive document ingestion supporting 12 formats including PDF, DOCX, PPTX, XLSX, Markdown, and Jupyter notebooks
02Hybrid search fusing LanceDB BM25 keyword matching with multilingual dense vector similarity via Reciprocal Rank Fusion
03Local-first architecture ensuring indexing, querying, and embeddings remain strictly on the host machine
04Support for multi-index serving, enabling one server instance to manage isolated indexes across projects
05Optional local OCR and classification extras utilizing Tesseract and Apple Vision without cloud APIs
0660 GitHub stars
ユースケース
01Grounding local AI coding assistants (such as Claude Code or Cursor) with context from internal project documentation and codebases
02Extracting answers and specific passages from large collections of technical manuals, PDFs, and spreadsheets locally
03Performing precise full-text and semantic queries across multi-language document archives without recurring API costs