qmd

qmd - Quick Markdown Search

Local search engine for Markdown notes, docs, and knowledge bases. Index once, search fast.

When to use (trigger phrases)

• "search my notes / docs / knowledge base"
• "find related notes"
• "retrieve a markdown document from my collection"
• "search local markdown files"

Default behavior (important)

• Prefer qmd search (BM25). It's typically instant and should be the default.
• Use qmd vsearch only when keyword search fails and you need semantic similarity (can be very slow on a cold start).
• Avoid qmd query unless the user explicitly wants the highest quality hybrid results and can tolerate long runtimes/timeouts.

Prerequisites

• Bun >= 1.0.0
• macOS: brew install sqlite (SQLite extensions)
• Ensure PATH includes: $HOME/.bun/bin

Install Bun (macOS): brew install oven-sh/bun/bun

Install

bun install -g https://github.com/tobi/qmd

Setup

qmd collection add /path/to/notes --name notes --mask "**/*.md"
qmd context add qmd://notes "Description of this collection"  # optional
qmd embed  # one-time to enable vector + hybrid search

What it indexes

• Intended for Markdown collections (commonly **/*.md).
• In our testing, "messy" Markdown is fine: chunking is content-based (roughly a few hundred tokens per chunk), not strict heading/structure based.
• Not a replacement for code search; use code search tools for repositories/source trees.

Search modes

• qmd search (default): fast keyword match (BM25)
• qmd vsearch (last resort): semantic similarity (vector). Often slow due to local LLM work before the vector lookup.
• qmd query (generally skip): hybrid search + LLM reranking. Often slower than vsearch and may timeout.

Performance notes

• qmd search is typically instant.
• qmd vsearch can be ~1 minute on some machines because query expansion may load a local model (e.g., Qwen3-1.7B) into memory per run; the vector lookup itself is usually fast.
• qmd query adds LLM reranking on top of vsearch, so it can be even slower and less reliable for interactive use.
• If you need repeated semantic searches, consider keeping the process/model warm (e.g., a long-lived qmd/MCP server mode if available in your setup) rather than invoking a cold-start LLM each time.