Layer1 Doc Search

Doc Search

Use this skill for read-only document retrieval from the Doc Search source namespace.

Do not use this skill for structured market prices, financial time series, trading-strategy backtests, direct recommendations, or future event schedules. Descriptive event-level backtest tools in this skill are allowed only for retrieved historical event evidence.

Runtime Binding

This skill defines the Doc Search logical source contract; it does not declare a separate user-visible MCP connection.

• In WorkBuddy, call Doc Search tools through the single tongzhou-fin-research Connector. Runtime tool names use doc_search__<tool>, for example doc_search__search_announcements.
• In Codex and retained legacy clients, the same canonical tool may appear as a bare name or a client-qualified server/tool name. Match the canonical tool suffix and keep the Doc Search source label.
• Do not bypass an available tongzhou-fin-research Connector by selecting similarly named global search tools such as search_hot_news or search_company_news.

Progressive References

Load only the reference needed for the current question:

• references/news.md: general news, company news, hot news, morning briefings
• references/announcements-events.md: announcements, structured events, normalized events, entity timelines
• references/research.md: company research, industry research, strategy research, broker views
• references/document-detail.md: single/bounded-batch detail retrieval and source coverage statistics
• references/examples.md: cross-domain and cross-MCP few-shot routing examples
• references/limitations.md: source coverage, realtime wording, empty results, internal-field boundaries

Read this skill's references/<name>.md file before filling detailed tool parameters.

For combination queries, read multiple references. Example: "今天新能源为什么大跌" may need references/news.md, references/announcements-events.md, and layer1-fin-data/references/screening.md.

Routing

Choose the narrowest retrieval family first:

The table uses canonical tool suffixes for readability. Under the WorkBuddy Connector, prepend doc_search__ when selecting the actual tool.

Required Rules

• Prefer specific tools over search_documents when the content type is clear.
• 空召回不是公司事实：只有业务调用成功且目标列表为空时才标记 empty，并同时说明公司/主题、文档类型、来源范围和时间窗。
• 参数错误不是空召回。对可修正的 scope、时间窗或字段错误，先按工具合同修正；一个证据任务最多两次恢复，依次只做身份规范化、拆分文档类型、放宽一个过滤或使用同证据域的声明式 fallback。
• 非空但缺少标题、日期、来源或用户要求字段时标记 partial，保留已返回文档并列出缺口，不因一个字段缺失丢弃全部结果。
• 超时、服务失败、协议错误和授权要求分别标记 error / auth_required；不得把错误写成空召回，也不得切换到新闻、同舟观点或模型记忆伪装成公告/研报命中。
• Every search call must include at least one non-empty narrowing signal such as query, ticker, company, industry, content_type, source_name, time_window, or date range.
• Resolve relative time before calling tools: 今天/昨日/本周/本月/最近N天 must map to a supported time_window, days, or explicit start_date/end_date; include the actual date range in the answer.
• For "今天有什么新闻/热点", use search_hot_news(query="今日热点", time_window="1d", limit=10) instead of an empty query.
• For company research reports, pass the entity through company and/or ticker in search_research_reports; do not put a company name only in query when using a long window such as 1y.
• Each research call accepts one company/listing. Split explicit company/ticker lists into separately labeled calls; MULTI_ENTITY_RESEARCH_UNSUPPORTED is a parameter error, not an empty result.
• For a listing-specific research question, preserve the resolved ticker, market, and scope="listing"; do not silently remove ticker or replace listing results with issuer-title matches.
• If issuer-wide coverage is useful, make a separate scope="issuer" call and label it as a broader issuer result.
• Normal research calls use limit=5-10; the hard single-call maximum is 20.
• A listing search may return a short-lived continuation_ref. When more evidence is needed, make the next call with only that reference; the gateway restores the original filters. Continue only while continuation_status="available", and stop once enough evidence is collected or the status is complete, limit_reached, or unavailable. Treat limit_reached / unavailable as bounded retrieval rather than exhaustive coverage. Older responses may omit continuation_status; if they also omit the reference, stop without claiming completeness. Never display, decode, construct, share, or persist the reference.
• The configured announcement source is CNINFO-oriented and does not include native HKEX disclosures. A Hong Kong announcement empty result is a source-coverage gap, not evidence that the company made no disclosure.
• Return title, date/time, source, and short summary first. Do not fetch long body text unless the user asks for details.
• Treat snippet as search-match context, summary as a stored source summary, and content as bounded detail evidence. Never relabel one as another.
• In get_document, read content_kind before using content: research_viewpoint, body_excerpt, event_analysis, and transcript_excerpt have different evidence meanings. Type-specific fields are under metadata.
• summary_status="unavailable" means no stored summary was returned; error means the summary source failed. Do not fill either state from content or model memory.
• Use get_document only after a search result provides doc_id and a clear doc_type. For research results, treat doc_id as a short-lived opaque evidence reference: pass it back unchanged, never display or persist it, and rerun the search if it expires.
• Use get_document_summaries only for 2-5 explicitly selected search results needed by one evidence task. Pass each returned evidence_ref with its exact doc_type; never construct, persist, mix across users, or use batch summaries as an export path. This tool returns only stored summaries and must not return or reconstruct original text.
• Use get_document_source_coverage only when the user asks about source coverage, data time range, or whether an empty result may be a coverage gap. Treat its first/last dates as an aggregate snapshot, not proof of continuous daily coverage.
• Search results are candidates for selection. For research, call get_document only for 1-3 selected reports needed by the user's request or claim verification; never expand every result by default.
• Do not expose internal collection names, retrieval scores, request plans, or backend routing details to users.
• If no result is returned, say no matching documents were retrieved under the current filters. Do not invent events or reports.
• Event backtest tools return descriptive historical samples only. Do not present them as a future prediction, strategy backtest, target price, or recommendation.
• If a multi-part user question asks for document evidence plus another data domain, finish the document summary and then continue with the other domain before finalizing.
• Do not answer "will query / please wait" after tool use. Continue calling the needed tool or provide the completed answer.
• Treat a non-empty documents or events list as retrieved evidence. Summarize those items; do not call it an empty result.
• For coverage, freshness, timeline, and content-type caveats, read references/limitations.md.

Output Shape

Prefer compact lists or tables. Include:

• title
• entity/company/industry when applicable
• publish/release date
• source or broker
• short evidence summary
• limitation or data-source caveat when relevant