This study proposes a method for semantically enriching a building information model (BIM) by converting and supplementing material layer information in computer-aided-design (CAD) drawings into semantically rich BIM objects. Despite the widespread adoption of BIMs, practices relying on CAD drawings are still prevalent and fully automated conversion from CAD-to-BIM remains challenging. Previous studies on automatically converting 2D drawings into a BIM have primarily focused on geometric reconstruction, often overlooking textual data such as material layer annotations, which contain essential BIM object details. However, the recognition and conversion of material layer annotations to BIM objects pose several challenges, including inconsistent annotation formats, vague associations between annotations and objects, and missing material function and property information. To address these challenges, this paper introduces a four-step method: (1) semantic clustering of material layer annotations into groups corresponding to individual objects; (2) material layer function assignment using a newly proposed mistake-driven prompting technique, which incorporates feedback from common mistaken cases; (3) object type classification using a large language model based solely on material layer texts; and (4) automated generation of BIM objects semantically enriched with material property information in single-directional composite or light-framed structures. Validation on 90 real-world cases demonstrated a 99.9% weighted average adjusted Rand index for semantic clustering, a 99.5% weighted F1-score for layer function assignment, a 92.7% weighted F1-score for object type classification, and 100% accuracy in BIM object generation. This method is expected to enable a more complete 2D-to-BIM conversion by complementing geometric reconstruction with semantically rich material layer information.

This study aims to propose a large language model (LLM)-enhanced defect question-answering (QA) method that can secure private and sensitive data while yielding high performance. Prompt responses to residents’ complaints are crucial for preventing recurring defects. However, traditional defect analysis and response methods rely on the expertise of a few skilled workers, making it difficult to ensure timely responses. The rapid advancement of LLMs offers a potential solution for improving defect QA tasks. However, many companies prohibit the use of closed-source LLM services, such as ChatGPT, due to concerns about potential data breaches. One possible solution is to use open-source LLMs like Llama and BERT, which can be locally installed and used. However, open-source LLMs typically perform worse than closed-source LLMs. Although the performance of open-source LLMs can be greatly improved through fine-tuning, the preparation of training datasets requires a significant amount of time and labor. To address these challenges, this study proposes a hybrid defect QA method that deploys an open-source LLM for defect management to secure sensitive information, and a closed-source LLM for generating a training dataset to reduce both the time and labor required. To validate the proposed method, we compare it to the state-of-the-art LLMs, GPT-4o and Llama 3, as well as graph retrieval-augmented generation (GraphRAG)-based QA systems, which have been extensively studied recently. Our results show that the hybrid LLM-based QA method achieved the highest ROUGE score of 81.6%. These findings demonstrate superior practical applicability, enabling cost-effective data generation and reliable domain adaptation within a secure data environment. This approach is beneficial for domain-specific tasks beyond defect management, where the accurate provision of specialized information and integration of historical knowledge are essential.