qmd is a local search engine for markdown files, recommended in the Karpathy gist as the tool to reach for when a bundle outgrows what index files can handle for navigation.

What it is

• Hybrid retrieval — BM25 keyword search plus vector similarity, with LLM re-ranking.
• On-device — runs locally; no data leaves the machine.
• Two interfaces — a CLI (an agent can shell out to it) and an MCP server (an agent can call it as a native tool).

When to use it in the LLM Wiki pattern

At small-to-moderate scale, the index file plus progressive disclosure is enough — do not add search prematurely. As the bundle grows to many hundreds or thousands of concepts, a search tool becomes the efficient way for the query operation to find candidate pages before reading them. Note the retrieved unit is still a synthesized concept, not a raw chunk — this is search over the wiki, not RAG over raw sources.

Citations

1. qmd — https://github.com/tobi/qmd
2. Karpathy — LLM Wiki gist