LLM-powered natural language interaction for voxel-based underground digital twin using task-aware retrieval and prompt engineering

초록

Efficient access to underground information is critical for infrastructure decision-making, yet existing systems require specialized technical expertise to query and interpret subsurface data. This paper proposes a modular task-aware RAG framework powered by LLMs, which dynamically selects the optimal retrieval strategy based on query intent. The framework is demonstrated on a road corridor and tunnel excavation, and evaluated across 400 queries and seven LLMs. Results demonstrate that no single retrieval strategy achieves optimal performance across all task types, confirming the necessity of task-aware retrieval. The task-aware query classifier achieves 100% routing accuracy on the primary dataset and 95.4% on 200 held-out validation queries. Task-specific prompt engineering delivers an average accuracy gain of 19.22% over generic prompting. Expert evaluation confirms strong usability and practical applicability. The proposed framework enables engineers and non-technical stakeholders to access, reason over, and visualize complex geotechnical information through natural language interaction with subsurface digital twins.

키워드