QMD

lastmod 2026-04-10created 2026-01-312 edits

Quick Markdown Search

About

A lightweight CLI search engine developed to search a variety of Markdown-based documents, such as personal notes, technical documents, and meeting minutes, locally
A hybrid search pipeline combining BM25 full-text search, vector semantic search, and LLM re-ranking
All computation runs locally, providing AI-level search quality without leaking personal information
Supports 3 search modes
- search: BM25-based keyword search
- vsearch: embedding-based semantic search
- query: the highest-quality mode, combining both approaches and re-ranking with an LLM
Built-in MCP (Model Context Protocol) server, so it integrates directly with LLM-based workflows such as Claude
Runs all models on-device via node-llama-cpp
- Automatically downloads and caches GGUF-format models such as embeddinggemma-300M, qwen3-reranker-0.6B, and Qwen3-1.7B
How search quality is improved
- Performs query expansion on the user's search request with the Qwen3-1.7B model
- Runs parallel searches through SQLite FTS5 and sqlite-vec
- Merges results with Reciprocal Rank Fusion (RRF)
- Re-evaluates document relevance with Qwen3-Reranker
- Balances accuracy and semantic similarity by adjusting weights per rank
Written in TypeScript and based on the Bun runtime. The index is stored in an SQLite database
MIT license
Open source by Tobi Lütke, founder and CEO of Shopify

author Zer0 sitemap