qmd

Install qmd via npm, add your notes directory with qmd collection add, run qmd embed to generate vectors, then use qmd vsearch for semantic queries or qmd query for hybrid search with LLM reranking. Everything runs locally with no cloud services.

qmd search is BM25 keyword-only search that runs in about 0.2 seconds with no models loaded. qmd vsearch uses vector embeddings for conceptual queries, and qmd query combines query expansion, BM25, vector search, and LLM reranking for the highest quality results.

Yes, qmd runs entirely on-device after the initial setup. The first run downloads about 2GB of local GGUF models for embeddings, reranking, and query expansion, after which all indexing and search happen locally with no cloud dependencies.

Run qmd mcp for stdio mode or qmd mcp --http --daemon for a persistent HTTP server on localhost:8181, then add the server to your agent's MCP configuration. This exposes tools like mcp_qmd_search, mcp_qmd_vsearch, and mcp_qmd_deep_search.

Cold start takes about 19 seconds because the three GGUF models must load into memory. Run qmd mcp --http --daemon to keep models warm for 2-3 second queries, or use qmd search for instant BM25-only lookups that need no models.

qmd requires Node.js version 22 or higher and runs on macOS and Linux. On macOS you also need Homebrew's SQLite because the system SQLite lacks extension loading support.