qmd Search

**Quick judgment** Solid, mature local search tool for markdown, notes, and code that is well-suited as an OpenClaw retrieval/backbone skill, especially in privacy-sensitive or offline environments. It offers strong functionality (BM25 + vector + rerank) and good performance, but requires careful scoping and operational setup to avoid leaking sensitive local content. **Recommended scenarios** - Use as the primary mechanism for an agent to **find relevant files, code, configs, and docs** across large local projects. - Prefer over raw `find` when you need **content-based search, semantic understanding, or reranked results** instead of simple filename/grep-style matching. - Use collection-based workflows (`qmd collection add/list`, `qmd update`) to create **project-specific knowledge bases** that the agent can reliably query. - Use `qmd search --files` for **fast file-path discovery**, `qmd search --full` or `qmd get` for **content extraction**, and `qmd query` when you want **best-quality combined retrieval**. **Key risks / limitations** - **Data exposure risk:** Because it searches local files, an agent using this skill can surface sensitive or private content if collections are not carefully defined and prompts are not constrained. Strongly recommended to: - Restrict which directories are added as collections. - Avoid indexing secrets or credentials, or apply masking policies. - Combine with higher-level guardrails on what content the agent is allowed to return. - **Setup/maintenance overhead:** To get the best results, someone must: - Define collections and masks (e.g., `--mask "*.md,*.py"`). - Run `qmd update` after file changes. - Run `qmd embed` to enable vector search (initial one-time cost, may take minutes). If an agent is expected to manage this itself, prompts must clearly describe when and how to run these commands. - **Scope awareness:** Search quality and safety both depend on **proper collection scoping** (`-c name`). If an agent forgets to specify collections, results may be noisy, incomplete, or inadvertently broad. **Overall** A strong choice as a **local retrieval and file-discovery skill** for OpenClaw agents operating on a developer’s machine or project workspace. Best used in environments where local indexing can be carefully curated, with explicit instructions to the agent on using collections, updating indices, and limiting exposure of sensitive content.