schema-bound-document-attribute-extraction

# Skill: schema-bound document attribute extraction ## 1. Capability Definition & Real Case * **Professional Definition**: The ability to read unstructured documents and map their contents onto a predefined schema, label space, or attribute inventory, preserving the document-grounded evidence for each extracted field. The challenge is not just spotting keywords, but determining whether the source text genuinely expresses the target label, attribute, severity, or class under a consistent annotation policy. * **Dimension Hierarchy**: Document Validation & Structuring->Extraction->schema-bound document attribute extraction ### Real Case **[Case 1]** * **Initial Environment**: A clinical note sentence is reviewed under a schema of social-determinants-of-health categories and adverse-status definitions. * **Real Question**: "Classify the sentence under the SDoH schema: Pt missed appointment because her sister couldn't drive her today." * **Real Answer**: Transportation issue; adverse SDoH mention. * **Why this demonstrates the capability**: The sentence does not merely mention travel in the abstract; it indicates a transportation barrier that affects care access. The correct output therefore depends on applying the schema, not on broad semantic similarity alone. This is a classic example of schema-bound sentence labeling grounded in document evidence. --- **[Case 2]** * **Initial Environment**: A clinical note sentence is classified under the same schema, but this time the mention is relational and non-adverse. * **Real Question**: "Classify the sentence under the SDoH schema: Pt and her husband came into my office today." * **Real Answer**: Relationship mention; not adverse. * **Why this demonstrates the capability**: The sentence contains evidence for the relationship category, but it does not indicate a support need or adverse condition. The task therefore tests whether the agent can separate category detection from severity or adverse-status detection. That separation is central to reliable structured extraction. --- **[Case 3]** * **Initial Environment**: A financial or legal document is processed under a fixed task schema such as named-entity recognition, key-information extraction, or categorical classification. * **Real Question**: "Extract the target field or label only if the document evidence satisfies the schema definition." * **Why this demonstrates the capability**: This class of benchmark evaluates whether the model can obey annotation guidelines and output normalized fields rather than loose free-form descriptions. The important capability is consistency of evidence-to-label mapping across many documents. That makes it directly relevant to document agents that annotate, review, or populate structured forms. ## Pipeline Execution Instructions To synthesize data for this capability, you must strictly follow a 3-phase pipeline. **Do not hallucinate steps.** Read the corresponding reference file for each phase sequentially: 1. **Phase 1: Environment Exploration** Read the exploration guidelines to discover raw knowledge seeds: `references/EXPLORATION.md` 2. **Phase 2: Trajectory Selection** Once Phase 1 is complete, read the selection criteria to evaluate the trajectory: `references/SELECTION.md` 3. **Phase 3: Data Synthesis** Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data: `references/SYNTHESIS.md`