qmd

You can search local markdown notes and meeting transcripts entirely on-device using a hybrid retrieval pipeline that combines BM25 keyword search, vector embeddings, and LLM reranking. This keeps your personal knowledge base private and runs without cloud services.

Hybrid retrieval merges BM25 keyword search with semantic vector search and then applies LLM-powered reranking to refine relevance. This combination ensures both exact keyword matches and contextual meaning are captured when querying local knowledge bases.

Yes, you can query project documentation through MCP integration for agent-based workflows or use the CLI for ad-hoc queries. The daemonized MCP HTTP mode supports low-latency queries for targeted information retrieval.

You need Node.js version 22 or higher and the qmd installation to run the search pipeline. Optional GGUF local models can be configured for generating vector embeddings and performing LLM reranking on your machine.

The best way is to query your meeting transcript collection using the hybrid search pipeline. It will retrieve and return the top-ranked passages containing design decisions, which you can then use for summarization or follow-up tasks.

Limitations depend on your local hardware capacity when running optional GGUF models for embeddings and reranking. Processing large project documentation collections on-device may be constrained by available memory and compute resources.