Bidding RAG Knowledge Retrieval Assistant

Problem it addresses

Tender and procurement search often has three engineering pain points: announcements are fragmented and noisy; users describe needs in natural language, such as “AI large-model related tender projects in the last 30 days nationwide”; downstream RAG and Agent systems need structured, traceable facts rather than unsupported summaries. The Bidding RAG Knowledge Retrieval Assistant turns natural-language queries into executable bid-search conditions and supports reviewable answers around announcement text, structured fields, and source links.

How it works

The skill does not add new data endpoints. It reuses a limited whitelist of Shibo Technology bidding-data APIs. A typical flow includes:
- Condition extraction: extract keywords, region, industry, purchaser, winner, project ID, amount range, time range, and exclusion terms.
- Search and summarization: use SearchProjectForAI for list results, getZTBProjectDetail for announcement body, and getZTBStructreDetail for structured fields.
- Traceability: use getCollectUrl to retrieve the original announcement URL, so amount, date, project ID, and other fields can be verified against the source.
- Condition rewriting: when the request is vague, aiSearchSubmitPolling can rewrite natural-language needs into searchable conditions.

Default query rules reduce ambiguity: if no region is given, assume nationwide; if no time is given, search queries default to the last 30 days, while trend or profile tasks may extend to 90 days or 12 months with the scope explained. Results filter out clearly irrelevant items such as recruitment, exams, training, meetings, news, and personnel announcements, and prioritize projects with high relevance, recent publication, clear amount, clear purchaser, and attachment or source link.

Boundaries

The skill is suitable for bid search, announcement summarization, structured parsing, opportunity screening, and RAG data access, not for promising bid success, bypassing procurement rules, or replacing formal procurement review. It depends on BBIAO_API_KEY; if the key is missing, the workflow should stop and ask the user to obtain it. API-returned fields must be treated as facts, and analytical recommendations must be separated from those facts. When results are empty, the response should explain the query scope and suggest adjustable conditions instead of inventing project names, amounts, dates, or source links.