Design and Evaluation of StructCPT: A Retrieval-Augmented Generation System Specialized for the Korean Architectural Structural Engineering Domain

Recent advancements in the field of architectural structural engineering have increasingly integrated artificial intelligence (AI) technologies that employ large language models (LLMs). However, these technologies face notable limitations, such as failing to incorporate region-specific structural design standards (e.g., KDS) or producing inaccurate and hallucinatory information. Retrieval-augmented generation (RAG) has emerged as a viable approach to overcoming these challenges. This study proposes StructCPT, an RAG system tailored for the Korean architectural structural engineering domain, and evaluates its performance. StructCPT is a domain-adapted retriever that extracts pertinent information from a Korean-language structural engineering knowledge base in real-time. It leverages maximum similarity retrieval (MAXIM), a contrastive learning-based embedding technique designed to enhance semantic similarity between queries and documents. Experimental results show that StructCPT consistently and significantly outperforms conventional retrieval methods, including BM25, Contriever, and SPECTER, across key quantitative metrics (Recall@10, MRR, and Precision@10). Notably, substantial improvements were observed in handling specialized structural engineering terminology and syntactically complex queries, resulting in enhanced retrieval accuracy and recall. Furthermore, StructCPT demonstrated practical scalability in terms of retrieval speed and memory usage, confirming its applicability to real-world scenarios. This study represents the first implementation of an RAG system specifically customized for structural engineering. It also outlines future developments, including the integration of multimodal data (e.g., drawings, images) and the incorporation of knowledge graphs, thereby laying a foundational basis for safe and trustworthy AI-driven decision support in structural engineering.