| name | complex document layout reasoning |
| description | Use this skill when the user wants questions about flowcharts, diagrams, or info hidden in page headers, footers, and complex 2D structures. Trigger it for requests like 'follow the process in this diagram', 'read the chart layout', 'what comes after this step in the map', or 'check the footer for the document code'. This is for scenarios where the answer isn't just in a text block but depends on where things are placed on the page. |
Skill: complex document layout reasoning
1. Capability Definition & Real Case
- Professional Definition: The ability to resolve queries by interpreting visual-spatial relationships and structural hierarchies within visually rich documents (VRDs), specifically targeting 2D layouts like flowcharts, nested diagrams, and positional metadata. This involves multi-hop reasoning where the 'hop' is defined by following a structural link (e.g., an arrow in a flowchart) rather than just a semantic bridge.
- Dimension Hierarchy: Multi-Source Evidence Composition->Cross-Document Synthesis->complex document layout reasoning
Real Case
[Case 1]
- Initial Environment: A bounded corpus of industrial process slides. One specific slide contains a complex '2D Layout' flowchart detailing the chemical pretreatment of biomass materials.
- Real Question: What is the temperature of Steam explosion of Pretreatment for Switchgrass and Sugarcane bagasse preparation?
- Real Answer: 195-205 Centigrade
- Why this demonstrates the capability: To answer this, the agent must first identify the 'Pretreatment' block in a complex diagram, then follow the specific pathway for 'Steam explosion' to find the associated numeric temperature value. This demonstrates the ability to reason over spatial connections (2D layout) rather than simple text retrieval.
[Case 2]
- Initial Environment: A multi-page document collection regarding textile manufacturing standards. One page contains a directional process map representing automated 'Weft Preparation'.
- Real Question: What are the steps of Weft Preparation between Spinning bobbin and Weaving?
- Real Answer: The intermediate steps between these two nodes in the process flowchart.
- Why this demonstrates the capability: The answer is not a single sentence but an entry in a sequence. The agent must locate the 'Spinning bobbin' start node and the 'Weaving' end node and extract the labels from all nodes positioned between them in the layout, demonstrating sequential reasoning over a visual structure.
[Case 3]
- Initial Environment: A large collection of quarterly financial reports (slides). The reports share similar textual content but differ in the specific data points rendered in 2D chart summaries.
- Real Question: What is the profit difference between the highest and lowest selling products in the 2024 reports?
- Real Trajectory: The agent scans the thumbnails of the entire corpus (Seeker role), identifies the 2024 quarterly summary slide, zooms in to inspect the high-resolution chart (Inspector role), and extracts the two values to calculate the difference.
- Real Answer: $24.7 Billion
- Why this demonstrates the capability: This involves a 'coarse-to-fine' approach where the agent must first find the correct visual layout (the 2024 chart) and then perform precise extraction from a 2D data visualization that text-only RAG would likely miss or misinterpret.
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:
-
Phase 1: Environment Exploration
Read the exploration guidelines to discover raw knowledge seeds:
references/EXPLORATION.md
-
Phase 2: Trajectory Selection
Once Phase 1 is complete, read the selection criteria to evaluate the trajectory:
references/SELECTION.md
-
Phase 3: Data Synthesis
Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data:
references/SYNTHESIS.md
