Geo-Intelligence: An Integrated Decision Support System for Geotechnical Investigation using Knowledge Graphs and LLMs

Abstract

This study presents “Geo-Intelligence”, an advanced decision-support system integrating knowledge graphs with Large Language Models (LLMs) to address the shortage of skilled geotechnical engineers. To overcome the limitations of conventional vector-based Retrieval-Augmented Generation (RAG) in handling spatial and geometric contexts, we propose a pseudo-3D spatial search model utilizing dynamic virtual edges. The system structures diverse geotechnical data format (e.g., XML, CSV, and JSON) into an ontology comprising geospatial and stratigraphic layers. Furthermore, it implements a semantic integration algorithm that dynamically synchronizes unstructured geological observation texts with core imagery based on depth-interval logic. Quantitative evaluations using practical queries reveal that the proposed GraphRAG approach significantly outperforms conventional RAG methods in accuracy. Additionally, the implementation of “Search Guardrails” strictly mitigates hallucinations, ensuring deterministic reliability and reproducibility of the generated answers. Accessible via a Web-GIS interface, this framework provides highly reliable, data-driven insights for site-specific risks and equipment optimization. Ultimately, it establishes a robust foundation for standardizing expert knowledge in geotechnical engineering.