跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
其他技能

personal-document-search

Use when the user wants to find or ask questions about their local files, cloud documents, mail, or Agent conversation history across Local Files, Agent Conversations, Feishu, official Notion MCP, Google Drive, and Mail. Rewrite queries for each source, search independently in parallel, read evidence on demand, and synthesize a cited answer.

LazyAGI · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add LazyAGI/LazyMind --skill personal-document-search

技能说明


name: personal-document-search description: Use when the user wants to find or ask questions about their local files, cloud documents, mail, or Agent conversation history across Local Files, Agent Conversations, Feishu, official Notion MCP, Google Drive, and Mail. Rewrite queries for each source, search independently in parallel, read evidence on demand, and synthesize a cited answer.

Personal Document Search · 个人文档检索

Scheduling rule

Before each tool round, collect the ready, independent searches and capability discovery calls for the selected Sources. Submit them together as multiple native tool_calls in one assistant response, using the host's supported concurrency. Build each batch in this order: first the ready initial searches or next discovery prerequisites across Sources, then evidence reads or refinements. If another selected Source has an unstarted available search or discovery call, include it before local file reads or repository inspection. A batch containing only local inspection is incomplete while independent Source discovery/search calls are ready. Finding a promising local project does not remove the other selected Sources. Describe calls as parallel only when they are actually dispatched together and the host supports concurrent execution; otherwise report the execution limitation.

Find information in the user's personal document space, including cloud documents, mail and saved Agent conversations. A Source is a logical information domain, not its backend: Agent Conversations remain separate from Local Files even when both use filesystem tools. Direct search and reading of raw Agent history files is allowed; a dedicated catalog, session reader or export is not required. Web, Wikipedia, academic search and external databases are outside this skill; an Agent's local session database is only a history storage format.

0. Reuse document directories when evidence is sufficient

Call list_document_directories first, then open_document_directory for matching Intent. These references are discovery context, not proof: read the original sources through existing authorized tools. Only omit a fresh cross-source search when that evidence supports the question's scope, time and completeness needs. Current-state, completeness, conflicts and missing evidence require actual expanded search. Explicit source limits always win. A directory mention is not an exclusive scope unless the user says so. Catalog listing does not count as directory use.

Do not organize task results yourself or add an extra recording/classification call. The runtime captures executed searches/reads and the idle organizer consumes them. Only use directory mutation tools for an explicit user request. Pinned entries have fixed positions; directory pin controls retention. A different organization creates a new directory rather than rewriting the original structure.

1. Select sources and discover capabilities

Extract topic, entities, people, time range, filename, project and source hints. Respect explicit scope. Strong hints such as “昨天张三发我的邮件” can select Mail alone. A project or topic name is a query hint, not a restriction to Local Files; finding its repository does not establish that cloud documents, mail or conversations are outside the request. When directory evidence is insufficient and there is no explicit or strong source hint, attempt all six supported Sources: Local Files, Agent Conversations, Feishu, Notion, Google Drive and Mail. An exact path or URL can go directly to its reader without an unnecessary discovery search.

Read source-routing.md for the selected Sources. Track each selected Source separately through an executed search, an actual access failure, or an evidenced capability gap. Agent Conversations are not covered by searching the project repository or recalling conversation memory. Use the existing filesystem tools and authorized history roots described in source-routing.md; absence of a dedicated conversation search tool is not itself a capability gap. Before answering, reconcile this coverage against actual tool calls. If a selected Source remains unattempted and is available, execute its search before concluding. Inspect actual registered tools and schemas; expand the existing capability groups as needed. Attempt the configuration/authorization entry for an unconfigured Source, then continue independent work. Do not silently omit it or invent an entrypoint. If no capability is exposed, record that limitation. Only say a configuration card was offered when the runtime returned evidence of it. Do not repeat pending calls.

2. Rewrite and search in parallel

Read query-rewrite.md. For each selected Source, adapt one information need into appropriate query text, filters and scope; do not copy the full question to every tool or multiply it into unrelated subquestions.

Apply the scheduling rule to discovery as well as search. For example, when these capabilities are exposed and their Sources are selected:

One response: local search + cloud-tool discovery + mail-tool discovery + Notion capability discovery
Later responses: expand newly discovered providers; batch their ready independent searches

Order calls only where one needs another's returned schema, locator or result: cloud-tool discovery precedes its provider discovery, which precedes that provider's search. These dependencies do not block work in other Sources. Do not wait for all Sources to become ready together. A failure, zero result or pending authorization in one Source must not stop the others. Independently accessible Agent corpora can also run in parallel within the Agent Conversations Source.

3. Read evidence on demand

For Agent history, use this sequence: hit → original path/physical line → message body → answer. Once a useful path and line are visible, read that location; do not finish paging the entire search r

阅读完整来源 ↗