Knowledge generation and evaluation for construction quality assessment using retrieval-augmented large language models

To address the problems of delayed knowledge updating, inefficient information acquisition, and difficulty in cross-project knowledge reuse in civil engineering construction quality assessment, this paper proposes a retrieval-augmented generation (RAG)-based methodology for injecting domain knowledge into large language models (LLMs). Multi-source knowledge, including national standards, journal articles, textbooks, and engineering reports, was systematically collected, cleaned, semantically segmented, and encoded using the bge-large-zh-v1.5 model to build a Milvus vector database. A general-purpose LLM was selected via semantic perplexity, and similarity-based retrieval was employed to construct prompt templates integrating role settings, prior knowledge, user queries, and response requirements, thereby enabling domain adaptation without additional training. An evaluation dataset containing 390 objective questions from registered architect and supervising engineer qualification examinations was constructed, and six representative LLMs along with their RAG-enhanced counterparts were evaluated. The results show that the average accuracy of general LLMs is 61.3%, with DeepSeek achieving the highest at 73.9%; after RAG enhancement, the accuracy rises to 86.3%, with significant improvements across all subjects. The proposed method effectively mitigates the hallucination phenomenon and provides traceable answers with source citations, enhancing both professionalism and interpretability, and offering a viable technical pathway for the digital transformation of the civil engineering industry.