VIS Publications Dataset Report
Generated autonomously for the 2025 IEEE VISxGenAI Workshop Challenge, this report demonstrates our multi‑agent system's dataset‑agnostic approach. The system handled the complete workflow independently:
- Data Understanding: The Field Refiner cleaned data and inferred column semantics, while the Dataset Describer and Field Expander (with web search) created semantic schemas and resolved cryptic codes
- Analysis: The Insight Planner operated within a ReAct loop, using statistical profiles to ground planning, while the Dataset Deriver crafted and repaired DuckDB queries
- Visualization: Rather than token‑intensive LLM‑based chart design, the Dataset Visualizer used Draco's rule‑based solver for principled visualization recommendations
- Reporting: The Report Narrator generated textual descriptions using vision‑language models, assembled into this Observable Notebook 2.0 report
The report combines AI‑authored insights with reader‑driven exploration via Mosaic and Quak for cross‑filtering and data interaction. Each insight includes an executable Marimo notebook for full traceability, while Langfuse captures detailed execution traces for transparency.
Source: VisPub Dataset