Motivation
Conventional AIS analysis requires SQL/GIS/Python expertise, while single-prompt LLM agents cannot balance factual accuracy and analytical reasoning.
Methodology
A two-stage agent: first classify query types (fact/aggregation vs. inference/analysis), then apply tailored prompts—compact for accuracy or CoT for reasoning—executed via safe tool-calling on AIS datasets.
Contribution
- Hybrid prompt architecture validated on 100 AIS QA benchmark
- Balanced improvements in accuracy, reasoning, and efficiency
