DuckDB-RAG-MCP-Sample

This is a sample for embedding and vectorizing markdown documents so that they can be explained from MCP via RAG. Vectorization uses [Plamo-Embedding-1B](https://tech.preferred.jp/ja/blog/plamo-embedding-1b/). ## Features - Text extraction and vectorization from markdown files - Vector search using DuckDB - Persistence of vector data via Parquet files - Vector search from MCP ## Usage ### Generating Vector Data First, place the markdown files you want to search in a specific directory and convert them to Parquet files using the following command. ```bash uv run main.py --directory ~/path/to/markdown/files --parquet vectors.parquet ``` ### MCP Configuration #### Build The following command generates a single binary at `dist/server`. ``` uv run pyinstaller --clean --strip --noconfirm --onefile server.py ``` #### MCP Client Configuration Configure according to the client you want to use. For Claude Desktop, it looks like this: Please specify the file you converted earlier for VECTOR_PARQUET. ```bash uv run mcp install server.py -v VECTOR_PARQUET=/path/to/vectors.parquet ``` It will be configured as follows: ```JSON:~/Library/Application Support/Claude/claude_desktop_config.json { "mcpServers": { "DuckDB-RAG-MCP-Sample": { "command": "/path/to/dist/server", "env": { "VECTOR_PARQUET": "/path/to/vectors.parquet" } } } } ``` ### Starting Development Server ```bash uv run mcp dev server.py ``` ## License DuckDB RAG MCP Sample is provided under the Apache License, Version 2.0.