knowledge-rag
A local retrieval-augmented-generation server exposed over MCP that lets Claude Code, Cursor and other agent hosts search a project's documents through hybrid semantic and keyword search, with no external service or API key.
knowledge-rag is an MCP-native local RAG server built to be searched directly by Claude Code, Cursor, Windsurf, VS Code, Cline, Gemini CLI and Zed. It combines semantic and BM25 keyword search with cross-encoder reranking, parses more than 30 file formats, and includes operational features such as bearer authentication, Prometheus metrics, rate limiting, health probes, structured logging and zero-downtime reindexing.
Embedding runs in-process through FastEmbed's ONNX models, so no Docker, Ollama or separate embedding server is required beyond the first model download, and the system works offline afterward. It installs with pip and registers itself as an MCP tool that a supported agent can call directly.
The project positions itself against larger RAG frameworks such as LlamaIndex, LangChain and Haystack by offering an opinionated, fully local default rather than a large set of configurable building blocks.