qmd Local Search

**Judgment:** qmd Local Search is a solid, pragmatic local-search skill for markdown, notes, and code, well-suited as an upgraded alternative to `find`/`grep` when working with indexed collections. It combines traditional BM25 keyword search with vector semantics and LLM reranking, giving noticeably better relevance than simple text search while keeping everything on-device. **Strengths & Advantages** - Fully local: uses small local models (embedding, reranker, and generator) that are auto-downloaded; no API keys or network required. - High-quality search: supports BM25 (`search`), pure semantic (`vsearch`), and combined reranked queries (`query`) to balance speed and relevance. - Scalable for large trees: designed to avoid `find`/`grep` hangs on big directories by using pre-built collections and indexes. - Flexible output: can return just file paths (`--files`), structured JSON with snippets (`--json`), or Markdown (`--md`), which is ideal for tool and agent integration. - Good ergonomics for projects: collections, masks, and `qmd update` make it easy to maintain project- or vault-level indexes. **Limitations & Risks** - Requires one-time embedding/indexing per collection (`qmd embed`, `qmd update`), which adds setup cost and may be slow on very large corpora. - Search coverage is limited to defined collections; content outside those collections will not be found unless explicitly added and re-indexed. - As with any strong search over local files, it can surface sensitive or private documents very quickly; this amplifies existing local access rather than introducing new network risk, but it may surprise users who are not careful with what is indexed. - Quality of semantic search and reranking depends on the bundled small models; results may be weaker than cloud-scale models but are generally good for local workflows. **Recommended Scenarios** - Developers and power users needing fast, robust search over large codebases, config trees, or documentation directories. - Users maintaining markdown-based knowledge bases or note vaults who want semantic search without sending data off-device. - Agents or automations that must first discover relevant files/snippets locally (via JSON output) before performing analysis or answering questions. - Offline or privacy-sensitive environments where external API calls are not allowed, but advanced search capabilities are still needed.