Bidding Document Structuring & Project Discrimination LLM

Automated systems processing bidding documents typically encounter two concrete problems: First, the raw text is often interspersed with non-project content (e.g., corporate promotions, industry reports), leading to false triggers in downstream workflows. Second, even after confirming a document is project-related, accurately extracting fields like amounts and company roles from unstructured titles and bodies demands complex rules and manual verification.

Core Challenges and Workflow

This skill addresses these issues with a deterministic processing pipeline centered on progressive ambiguity reduction. It enforces the sequence of "discrimination, then classification, then structuring" to avoid forcibly processing non-project text.

1. Project Information Discrimination: First, the 项目信息甄别 interface is called to judge the input title and body. Only when the result is "Bid Information" or "Planned Construction Information" does the process advance. If it's "Non-Project Information" or "Abnormal Information", it stops immediately and outputs the reason, effectively isolating non-target text.
2. Classification and Structuring: For confirmed bid information, the 招中标分类推理 interface determines its specific category. Subsequently, the LLM Bid Information Structuring interface extracts fields. This interface supports custom rules (e.g., amount units, phone formats) and defaults to a minimal rule set if none are provided (e.g., fields without evidence return null, amounts default to RMB yuan).

Key Mechanisms and Validation

The skill's critical features are its rigorous validation and output specifications, ensuring results are machine-usable and traceable.
- Evidence-Chain Validation: Every structured field must have direct evidence in the source text or be explicitly marked as a model inference. Conversion rules, such as converting ten-thousand yuan to yuan, are strictly applied.
- Error and Status Handling: During batch processing, each item is handled independently; a single failure does not contaminate other results. The output must include 原始模型响应 (raw model response), 校验状态 (validation status), 缺失字段 (missing fields), and 冲突字段 (conflicting fields) to support quick human review.
- Interface Dependencies: Before invocation, the environment variable BBIAO_API_KEY must be configured; otherwise, the process terminates with a prompt on how to obtain it. The service address can be overridden via BBIAO_SERVER_URL, defaulting to https://gate.gov-bid.com.

Constraints and Important Notes

This skill is suitable for initial data cleaning and structured extraction from bidding texts, but users should note:
- It processes the "text" itself and does not guarantee the authority of its source.
- The output structured JSON fields may be incomplete (as indicated by 缺失字段), requiring handling by downstream systems or manual intervention.
- When there's a conflict between the classification interface result and the body's semantics, the skill retains both and marks them, meaning it favors presenting facts over forcing a single judgment.