A Hybrid Prompt Agent for Maritime AIS Data Analysis: Performance Optimization through Query Classification and Dynamic Routing

While Large Language Model (LLM)-powered agents are transforming conversational data analysis, their efficacy on specialized datasets, such as the Automatic Identification System (AIS), is highly sensitive to prompt engineering strategies. Conventional static prompting approaches fail to adequately address the broad spectrum of user query complexity, which ranges from straightforward fact retrieval to multi-step analytical reasoning. To mitigate this limitation, we propose a Hybrid Prompt Agent architecture. This system employs a preliminary query classification mechanism to dynamically select and apply an optimal prompting template: utilizing compact prompts for high factual precision and Chain-of-Thought (CoT) prompts for in-depth analytical tasks. Empirical evaluation on a custom-built benchmark of 100 diverse AIS-related queries demonstrates that our hybrid approach achieves an effective balance between quantitative accuracy and qualitative contextual relevance compared to single-prompt baselines. These findings validate dynamic prompting architectures as a practical, effective, and resource-efficient alternative to full model fine-tuning for developing specialized domain analysis tools.