Find your landing page message
Connect a product’s value to a specific audience and write a clear first-screen message.
Search and retrieve markdown documents from local knowledge bases using qmd. Supports BM25 keyword search, vector semantic search, and hybrid search with LLM re-ranking. Use for querying indexed notes, documentation, meeting transcripts, and any markdown-based knowledge. Requires qmd CLI installe...
--- name: qmd-3b82879b60c9 description: Search and retrieve markdown documents from local knowledge bases using qmd. Supports BM25 keyword search, vector semantic search, and hybrid search with LLM re-ranking. Use for querying indexed notes, documentation, meeting transcripts, and any markdown-based knowledge. Requires qmd CLI installed (bun install -g https://github.com/tobi/qmd). --- # QMD - Local Markdown Search Search and retrieve documents from locally indexed markdown knowledge bases. ## Installation ```bash bun install -g https://github.com/tobi/qmd ``` ## Setup ```bash # Add a collection qmd collection add ~/notes --name notes --mask "**/*.md" # Generate embeddings (required for vsearch/query) qmd embed ``` ## Usage Rules **Always use `--json` flag** for structured output when invoking qmd commands. ## Search Commands ### search (BM25 keyword search - fast) ```bash qmd search "authentication flow" --json qmd search "error handling" --json -n 10 qmd search "config" --json -c notes ``` ### vsearch (vector semantic search) ```bash qmd vsearch "how does login work" --json qmd vsearch "authentication best practices" --json -n 20 ``` ### query (hybrid with LLM re-ranking - best quality) ```bash qmd query "implementing user auth" --json qmd query "deployment process" --json --min-score 0.5 ``` ### Search Options | Option | Description | |--------|-------------| | `-n NUM` | Number of results (default: 5, or 20 with --json) | | `-c, --collection NAME` | Restrict to specific collection | | `--min-score NUM` | Minimum score threshold | | `--full` | Return complete document content in results | | `--all` | Return all matches | ## Retrieval Commands ### get (single document) ```bash qmd get docs/guide.md --json qmd get "#a1b2c3" --json qmd get notes/meeting.md:50 -l 100 --json ``` ### multi-get (multiple documents) ```bash qmd multi-get "docs/*.md" --json qmd multi-get "api.md, guide.md, #abc123" --json qmd multi-get "notes/**/*.md" --json --max-bytes 20480 ``` ## Maintenance Commands ```bash qmd update # Re-index changed files qmd status # Check index health qmd collection list # List all collections ``` ## Search Mode Selection | Mode | Speed | Quality | Best For | |------|-------|---------|----------| | search | Fast | Good | Exact keywords, known terms | | vsearch | Medium | Better | Conceptual queries, synonyms | | query | Slow | Best | Complex questions, uncertain terms | **Performance note:** `vsearch` and `query` have ~1 minute cold start latency for vector initialization. Prefer `search` for interactive use. ## MCP Server qmd can run as an MCP server for direct integration: ```bash qmd mcp ``` Exposes tools: `qmd_search`, `qmd_vsearch`, `qmd_query`, `qmd_get`, `qmd_multi_get`, `qmd_status`--- name: qmd-3b82879b60c9 description: Search and retrieve markdown documents from local knowledge bases using qmd. Supports BM25 keyword search, vector semantic search, and hybrid search with LLM re-ranking. Use for querying indexed notes, documentation, meeting transcripts, and any markdown-based knowledge. Requires qmd CLI installed (bun install -g https://github.com/tobi/qmd). --- # QMD - Local Markdown Search Search and retrieve documents from locally indexed markdown knowledge bases. ## Installation ```bash bun install -g https://github.com/tobi/qmd ``` ## Setup ```bash # Add a collection qmd collection add ~/notes --name notes --mask "**/*.md" # Generate embeddings (required for vsearch/query) qmd embed ``` ## Usage Rules **Always use `--json` flag** for structured output when invoking qmd commands. ## Search Commands ### search (BM25 keyword search - fast) ```bash qmd search "authentication flow" --json qmd search "error handling" --json -n 10 qmd search "config" --json -c notes ``` ### vsearch (vector semantic search) ```bash qmd vsearch "how does login work" --json qmd vsearch "authentication best practices" --json -n 20 ``` ### query (hybrid with LLM re-ranking - best quality) ```bash qmd query "implementing user auth" --json qmd query "deployment process" --json --min-score 0.5 ``` ### Search Options | Option | Description | |--------|-------------| | `-n NUM` | Number of results (default: 5, or 20 with --json) | | `-c, --collection NAME` | Restrict to specific collection | | `--min-score NUM` | Minimum score threshold | | `--full` | Return complete document content in results | | `--all` | Return all matches | ## Retrieval Commands ### get (single document) ```bash qmd get docs/guide.md --json qmd get "#a1b2c3" --json qmd get notes/meeting.md:50 -l 100 --json ``` ### multi-get (multiple documents) ```bash qmd multi-get "docs/*.md" --json qmd multi-get "api.md, guide.md, #abc123" --json qmd multi-get "notes/**/*.md" --json --max-bytes 20480 ``` ## Maintenance Commands ```bash qmd update # Re-index changed files qmd status # Check index health qmd collection list # List all collections ``` ## Search Mode Selection | Mode | Speed | Quality | Best For | |------|-------|---------|----------| | search | Fast | Good | Exact keywords, known terms | | vsearch | Medium | Better | Conceptual queries, synonyms | | query | Slow | Best | Complex questions, uncertain terms | **Performance note:** `vsearch` and `query` have ~1 minute cold start latency for vector initialization. Prefer `search` for interactive use. ## MCP Server qmd can run as an MCP server for direct integration: ```bash qmd mcp ``` Exposes tools: `qmd_search`, `qmd_vsearch`, `qmd_query`, `qmd_get`, `qmd_multi_get`, `qmd_status`
Imported from the listed AgentSkill.sh category. Complete supplied package files and notices are retained. Required external tools and credentials are not installed. No bundled code was executed during import. Individual agent task behavior has not been tested.