Bidding RAG Data Assistant

Core Challenges in Processing Bidding Data for RAG

When building a question-answering, retrieval, or recommendation system for the bidding sector, developers face a concrete problem: raw bid announcements are messy and scattered. Critical information like the buyer, amount, qualification requirements, and attachments is often buried in lengthy text or PDF files. Manually extracting and structuring this data for RAG ingestion is not only time-consuming but also makes consistent, traceable sourcing difficult. This leads to unreliable citations in subsequent AI-powered Q&A.

Skill Workflow: From Natural Language Queries to Structured Data

This skill automates the process by encapsulating the professional data APIs of Shibo Technology. Its core capabilities and key steps are:

• Intelligent Intent Recognition & Parameter Completion: Users can ask in natural language, e.g., "find government procurement winning bids in Beijing over the last three years." The skill parses the intent, automatically fills in default parameters (like default time range, region codes), and generates the corresponding API call.
• Core Data Interface Invocation: The skill primarily uses a dedicated whitelist of interfaces:
  • SearchProjectForAI: For natural language searches to get a list of matching projects.
  • getZTBProjectDetail: Retrieves the original text body of the announcement.
  • getZTBStructreDetail: Fetches 78 structured fields such as amount, contact information, and winning bidder, which is key to making data analyzable.
  • getZTBProjectFiles & getCollectUrl: Used to trace the list of attachments and the original source URL, ensuring data verifiability.
• RAG-Focused Data Chunking and Standardization: After data retrieval, the skill guides semantic chunking. For example, a long announcement is split into semantic segments like procurement requirements, qualification conditions, and scoring methods. The output uniformly includes metadata such as title, structured fields, publication date, region, announcement type, and original link, preparing it for ingestion.

Applicable Scope and Important Considerations

This skill focuses on solving the "retrieval-extraction-structuring" problem for bidding data. The following points should be noted:

1. Data Dependency: All functions rely on Shibo Technology's API services and a valid API Key (configured via the environment variable BBIAO_API_KEY). The skill itself does not cache data; it calls interfaces in real-time.
2. Content Scope: The skill processes core procurement opportunity announcements like bids, winning results, contracts, and planned constructions. It correctly excludes non-target content such as recruitment, training, and conference news.
3. Output Characteristics: The skill outputs structured, traceable "data tuples," not generated answers. For missing fields, it explicitly marks them as "not disclosed in the announcement" and does not fabricate information.
4. Quota Management: Invocations consume API quotas. When the balance is insufficient, the skill provides a fixed link to obtain service packages and stops the current task. It waits for the user to manually confirm a top-up before retrying, avoiding automatic billing or infinite loops.
5. Security Boundary: API Key storage and handling follow security rules. It is prohibited from being leaked in any output and supports reading from the platform's secure credential store.