QMD is an on-device search engine designed for personal knowledge management. With 13.8k GitHub stars, it indexes markdown notes, meeting transcripts, documentation, and knowledge bases, enabling both keyword and semantic search entirely locally without sending data to external servers. ## The Local-First Search Problem Developers and knowledge workers accumulate vast collections of notes, documentation, and transcripts across their machines. Traditional search tools either require cloud indexing, sacrificing privacy, or offer only basic filename matching that misses contextual connections. QMD bridges this gap by providing production-grade search capabilities that run entirely on your device. The project was created by Tobi Lutke (known as the CEO of Shopify), which signals both technical credibility and a focus on practical developer tooling. ## Three-Model Architecture QMD's search pipeline combines three local GGUF models running via node-llama-cpp, each handling a specific aspect of the search process: | Model | Purpose | Size | |-------|---------|------| | embedding-gemma-300M-Q8_0 | Vector embeddings | ~300MB | | qwen3-reranker-0.6b | Result re-ranking | ~640MB | | qmd-query-expansion-1.7B | Query enhancement | ~1.1GB | This approach delivers search quality comparable to cloud-based solutions while keeping everything local. The total model footprint of approximately 2GB is modest by modern standards. ### Hybrid Search Pipeline QMD combines BM25 full-text search via SQLite FTS5 with vector semantic search using local embeddings. Results are then re-ranked using a dedicated model and fused through Reciprocal Rank Fusion with position-aware score blending. This multi-stage pipeline preserves high-confidence exact matches while leveraging semantic understanding for conceptual queries. ### Query Expansion Before executing a search, QMD's query expansion model reformulates the user's query to capture related terms and concepts. Searching for 'React hooks performance' might expand to include 'useCallback', 'useMemo', 'rendering optimization', and related terms, dramatically improving recall without sacrificing precision. ## Integration Options ### CLI Tool The command-line interface provides standalone search for interactive use. Developers can index directories and run queries directly from the terminal. ### MCP Server QMD functions as a Model Context Protocol server, enabling AI agents like Claude Code or Codex to search your knowledge base as a tool. This transforms personal notes into a queryable resource that AI assistants can leverage during conversations. ### HTTP Transport For long-lived server instances, QMD supports HTTP transport, making it accessible to web applications and other services on the local network. ### Node.js/Bun SDK Programmatic access through the SDK allows developers to integrate QMD search into their own applications. ## Data Organization QMD supports multi-collection indexing with custom glob patterns. Each collection can target specific directories and file types. Hierarchical context tagging enhances search understanding by providing metadata about document relationships and categories. ## Practical Applications QMD excels at searching across meeting notes and transcripts to find past decisions and context. Technical documentation spanning multiple projects becomes instantly searchable. Research notes and paper annotations gain semantic connections that filename-based search cannot provide. Developer journals and daily logs become a queryable knowledge base. ## Limitations The initial indexing process can be slow for large document collections, as embedding generation requires processing each document through the local model. QMD currently focuses on text-based documents and does not handle images, PDFs, or binary formats natively. The 2GB model requirement may be significant on resource-constrained devices. The project is still in active development, with 123 open issues indicating ongoing refinement. ## Market Position QMD occupies a unique niche as a local-first semantic search engine. Alternatives like Khoj offer cloud connectivity, while tools like Obsidian provide search within their own ecosystem. QMD's advantage is its format-agnostic approach and MCP integration, which makes it a search backbone for any AI-powered workflow that needs access to local knowledge.