RAG-for-CR: A Retrieval-Augmented Generation Framework for Efficient, Accurate, and Traceable Querying of Construction Regulations

The effective retrieval and precise application of construction regulations are critical for the industry, yet face challenges such as low clause-matching accuracy, fragmented semantic context, and unclear answer provenance. Although retrieval-augmented generation (RAG) shows great potential, existing frameworks still struggle to adapt to the hierarchical structure of regulatory texts and lack systematic optimization capabilities. This study proposes a tailored RAG framework—RAG-for-CR (RAG for construction regulations)—to enhance the accuracy, interpretability, and practical utility of construction regulation queries. The framework consists of three core components: (1) a metadata-based decomposition method that preserves the hierarchical structure of clauses via structured annotation; (2) a dynamic multistrategy retrieval engine that integrates keyword search, semantic embedding, and metadata filtering, with adaptive weight adjustment based on user feedback; and (3) a traceable prompting system that ensures responses are supported by explicit regulatory clauses. A comprehensive evaluation demonstrates that RAG-for-CR performs exceptionally across diverse query tasks: it significantly outperforms baseline models in retrieval performance, improving hit rate (HR), recall, and mean average precision (MAP) by 20%, 29%, and 29%, respectively. In response quality, it ranks first in completeness (4.92), readability (4.83), and correctness (4.89). Ablation studies confirm the essential contributions of each component, and cross-lingual tests underscore its strong generalization capability. In practical assessments, the framework achieved an efficiency score of 4.13, with all performance indicators receiving positive feedback. These results highlight its practical value in meeting the construction industry’s need for efficient and accurate regulation retrieval and interpretation. Moreover, the framework exhibits transferability beyond construction, offering a scalable methodology for integrating domain-specific regulatory knowledge with large language models in other professional fields.