Journal of KIBIM ISSN:2288-1697(Print) 2288-1697(Online) 한국BIM학회논문집

Building Information Modeling (BIM) data, accumulated throughout the entire lifecycle of construction infrastructure, is a massive collection of information that includes not only geometric data but also complex topological and semantic relationships between objects. Recently, there has been an increasing number of attempts to explore construction data using natural language with Large Language Models (LLM); however, existing vector-based Retrieval-Augmented Generation (RAG) systems suffer from critical hallucinations because they fail to properly retrieve the object structures contained in Industry Foundation Classes (IFC) data. To overcome these limitations, this study introduces a Knowledge Graph-based RAG architecture and proposes a Hallucination Suppression Pipeline for BIM Graph RAG (NSP4BIM-RAG) that can robustly respond to diverse natural language user inputs (prompt variability), specifically for BIM Graph RAG. The proposed model controls the potential for Cypher query generation errors through a three-stage pipeline combining input normalization, multi-pattern few-shot prompting, and deterministic syntax validation and dynamic retry loops. As a result of performing harsh tests on a total of 21 query scenarios in Graph DB and local open-weighted LLM environments, it was confirmed that the proposed model significantly improved the accuracy of defense against topological relation inference and exception queries from 28.57% of the existing baseline to 90.48%, and significantly reduced the occurrence rate of fatal hallucinations such as false attributes and incorrect label generation.