QMD: Query Markup Documents

Mini CLI search engine for your docs, knowledge bases, meeting notes, whatever. All local, tracking current SOTA approaches.

Core Features

• Full-text search (BM25): Fast keyword-based retrieval
• Vector semantic search: Understanding meaning beyond keywords
• Hybrid search with re-ranking: Combines both plus LLM-powered relevance scoring

Architecture

1. Query Expansion: LLM generates variations for broader matching
2. Parallel Retrieval: Each variant searches both BM25 and vector indexes simultaneously
3. Ranking Fusion: Reciprocal Rank Fusion (RRF) with position-aware blending
4. LLM Re-ranking: Cross-encoder model scores final candidates with confidence metrics

Smart chunking preserves semantic units -- headings, code blocks, and sections stay together.

Technical Requirements

• Node.js >= 22 or Bun >= 1.0
• Three GGUF models auto-download on first use (~2GB total):
• EmbeddingGemma (300M) for vector embeddings
• Qwen3-Reranker (0.6B) for relevance scoring
• Custom QMD query-expansion model (1.7B)

Key Features

• Configurable collections from any directory with custom glob patterns
• "Context" layers that provide metadata to improve search relevance
• Multiple output formats (JSON, CSV, Markdown, XML) for agent integration
• MCP server for Claude and other LLM clients
• SDK for embedding in Node.js/Bun applications
• GPU acceleration automatic, can be overridden

Quick Start

npm install -g @tobilu/qmd
qmd collection add ~/notes --name notes
qmd embed
qmd query "search term"