qmdlisted

# qmd Local Markdown search engine: BM25 + vector + optional LLM reranking, all running on-device. This skill teaches Claude how to use it well — what mode to pick, what flags matter, what to avoid. ## Scope **Handles:** searching indexed Markdown collections, retrieving documents by path/docid/glob, registering new collections, refreshing the index. **Does NOT handle:** code search (use `grep`/`rg`), web search, structured DB queries, vector embeddings of non-Markdown formats. ## Install (one-time) ```bash # Bun (recommended by upstream) bun install -g https://github.com/tobi/qmd # OR npm npm install -g @tobilu/qmd ``` Verify: `qmd --version`. Requires `bun` (or Node ≥18) and `sqlite` (preinstalled on macOS). ## Mode selection — pick the cheapest mode that works | Mode | Speed | Model load? | Use for | |---|---|---|---| | `qmd search` | instant | none | **default** — keyword/BM25 lookup | | `qmd vsearch` | ~1 min cold | ~334MB embedding model | semantic similarity when keywords fail | | `qmd query` | slowest | +1.28GB rerank LLM | hybrid — only if user explicitly demands "best quality" and accepts wait | **Heuristic:** start with `qmd search`. If results are empty or noisy, escalate to `qmd vsearch`. Never start with `qmd query` — its multi-GB downloads surprise users on first run, and BM25 is sufficient for most collections under ~10k docs. ## Setup a new collection ```bash qmd collection add /path/to/markdown/tree --name <my-coll> --mask "**/*.md" qmd collecti