Enhancing Knowledge Management in Green Building Engineering: A Hybrid Retrieval-Augmented Generation Approach to Support Project Assessment

Green building assessment relies heavily on manual queries of complex engineering regulations, making the introduction of knowledge management theories crucial for improving assessment efficiency. However, the inherent characteristics of engineering regulations—dense terminology, complex clause dependencies, and frequent revisions—limit the effectiveness of traditional knowledge engineering. This study aims to explore the utilization of advanced information technologies to achieve reliable knowledge management for supporting engineering decision-making. From the perspective of acquisition, organization, retrieval, and presentation (AORP), a domain knowledge management framework is designed by integrating network analysis, information retrieval, and large language models (LLMs) with traditional knowledge management theories. A dual-structure knowledge network is proposed as the underlying carrier to characterize explicit rules and semantic dependencies, enabling lightweight knowledge organization. Furthermore, a hybrid retrieval algorithm capable of identifying and adaptively weighting domain terms was developed and embedded into a retrieval-augmented generation workflow. Facilitated by LLMs, the proposed framework supports more efficient, traceable, and scalable building assessment. This study covers regulations containing 225,604 words and validates the effectiveness of the proposed method using 132 practical test cases. The results showed that the proposed method achieves a recall rate of 0.959, effectively reduces generation hallucinations, and substantially improves the accuracy and reliability of knowledge presentation. More importantly, the experiments indicated that the performance gap among LLMs with different parameter scales narrowed after introducing domain knowledge, confirming that an effective knowledge management architecture in vertical domains has the potential to reduce dependence on large-parameter LLMs and provide a feasible pathway for low-cost deployment of intelligent applications in the industry.