kanna12580/kk-knowledge-agent

kk-knowledge-agent is a lightweight knowledge base system that lets you store documents and search them using natural language. You upload plain text content or .txt files, the system splits them into chunks and converts them into numerical embeddings using a local multilingual model (sentence-transformers), then stores those embeddings in ChromaDB. When you search, the system finds chunks that are semantically close to your query rather than just matching keywords, and streams the results back. The project has three parts. The backend is a FastAPI service that handles knowledge base management (create, list, update, delete), document ingestion, and search. It uses SQLite for document metadata and ChromaDB for the vector index. The frontend is a React and Vite demo app for browsing knowledge bases and uploading documents. The third part is an MCP (Model Context Protocol) server, which wraps the backend's search capability as a tool that AI agents can call directly. The README shows it being used with Claude Code and similar tools. To run it locally, you start the backend with uvicorn, the frontend with npm, and the MCP server separately with Python. The MCP config points the agent at the local server.py file. All three components communicate over HTTP on localhost. The README notes several planned improvements: Docker Compose for one-command startup, swapping SQLite for PostgreSQL and ChromaDB for Qdrant or Milvus at larger scale, adding PDF and DOCX upload, and adding hybrid BM25 and vector retrieval with reranking.