qmd

What This Skill Does

qmd is a high-performance, local-first hybrid search engine specifically engineered for Markdown-based knowledge bases and personal note collections. Unlike cloud-based indexing services, qmd keeps your sensitive data entirely local. It leverages BM25 (Best Matching 25) for instantaneous keyword-based document retrieval, ensuring that common searches are lightning-fast. For more complex, conceptual queries, it supports vector-based semantic search, allowing users to find content based on meaning rather than literal word matches. The skill integrates seamlessly with your local file system, providing an indexed environment that allows for efficient retrieval of specific document chunks or full-text results.

Installation

To get started, ensure your environment has Bun (>= 1.0.0) installed. On macOS, you must install the necessary SQLite dependencies by running brew install sqlite. Once prerequisites are met, install the tool globally via: bun install -g https://github.com/tobi/qmd. Post-installation, initialize your knowledge base by adding your directories with qmd collection add /path/to/notes --name notes. Finally, execute qmd embed to generate the vector index, which enables the skill's advanced semantic capabilities.

Use Cases

qmd is the ideal solution for users managing large volumes of local Markdown files, research papers, or daily journals. It is perfect for developers maintaining documentation in plain text, writers organizing manuscript drafts, and researchers who need a structured way to query their offline libraries. Use this skill when you know a specific keyword exists in a document, or when you are trying to remember a concept across multiple scattered files where exact wording might be uncertain.

Example Prompts

1. "Search my notes for all references regarding the Q3 product strategy update."
2. "Find related notes to my current project plan that discuss cloud infrastructure."
3. "Retrieve the full content of the markdown file about the local database configuration in my tech collection."

Tips & Limitations

Performance is prioritized by defaulting to keyword search (BM25). Use this as your primary mode for near-instant results. Reserve semantic searches (vsearch) only for when keyword matching proves insufficient, as this mode can introduce latency due to local model initialization. Avoid using the full hybrid query mode during interactive sessions, as LLM reranking can trigger significant timeouts. Always maintain your index by periodically running qmd embed after significant updates to your local files to ensure your search results remain accurate.