Digital twins are increasingly used in the Architecture, Engineering, and Construction (AEC) industry, but their adoption is often hindered by the need for specialised knowledge, such as database querying. This paper presents Graph-DT-GPT, a multi-agent framework that integrates Large Language Models (LLMs) with graph-based digital twins to enable natural language interaction. The framework is designed with modular agents, including decision, query generation, and answer extraction, and grounds all LLMs’ outputs in structured graph data to improve response reliability and reduce hallucinations. The framework is evaluated on two use cases: a city-level graph with over 40,000 building nodes and room-level apartment layout graphs. Graph-DT-GPT achieves 100% and 95.5% answer correctness using Claude Sonnet 4.5 and GPT-4o, respectively, in the city-scale case, and 100% correctness in the room-level case, significantly outperforming baseline methods including LangChain Neo4j pipelines by approximately 40% and 10%, respectively. These results demonstrate its scalability and potential to enhance accessible, accurate information retrieval in AEC digital twin applications. • Propose Graph-GT-GPT, an LLM-enabled multi-agent framework for graph-based digital twins. • Introduce modular agents for query decomposition, generation, and response synthesis. • Ground LLM outputs in graph data to reduce hallucinations and improve reliability. • Deploy prototypes that outperform the LangChain Neo4j toolbox and prompt-only baselines. • Handle complex reasoning tasks like shortest-path finding in room graphs.
Pan et al. (2026) studied this question.