A Localized RAG-LLM System for Safety Compliance Question–Answering in Working-At-Heights Construction
AI · 02.11.2026 · 7 (11) · 449
DOI: 10.3390/ai7110449 ↗Kurz gefasst
Fachartikel in der Zeitschrift AI, online erschienen am 02.11.2026. Verfasst von Chun-Ling Ho und 2 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Innovation Strategy of Fujian Provincial Science and Technology Projects.
Abstract (Originalsprache)
With the advancement of building digitalization and safety management, Working-at-Heights presents challenges due to its complex scenarios and dispersed regulations, making it difficult for on-site personnel to quickly and accurately query and assess compliance. While general-purpose large language models possess text capabilities, they lack professional knowledge adaptation, are not updated in a timely manner, and provide answers without sufficient evidence, thus failing to directly meet the safety and compliance requirements of Working-at-Heights. This paper deploys a localized retrieval-augmented generation (RAG) system based on RAGFlow and Docker. Addressing the safety consultation needs of Working-at-Heights, it transforms regulations and standards into a vector knowledge base. The system integrates keyword and semantic retrieval for precise positioning and aims to reduce hallucinations through source tracing and prompt constraints. Simultaneously, it connects to online retrieval to supplement the latest regulations, achieving a dual-track question–answering (Q&A) mechanism of “local priority, online supplementation.” A typical scenario test set of 100 questions was constructed to evaluate three system configurations across multiple dimensions, including accuracy, traceability, timeliness, and latency. The results show that faithfulness scores increased from 0.45 for the pure large model to 0.91 for local RAG and 0.97 for full-functionality integration, while context recall improved from 0.78 to 0.93 between the latter two configurations. However, response latency also increased from 2.1 s to 4.5 s and 6.8 s, respectively. These findings indicate that local RAG provides more reliable and traceable responses to basic compliance inquiries, whereas full-functionality integration offers superior performance in dynamic information supplementation and handling of unsafe requests. Overall, the integrated approach demonstrates potential for supporting on-site safety compliance consultation and risk assessment, providing a foundation for intelligent applications in construction safety management.
Förderung
- Innovation Strategy of Fujian Provincial Science and Technology Projects
Zitieren
APA
Ho, C., Chang, T., & Guo, X. (2026). A Localized RAG-LLM System for Safety Compliance Question–Answering in Working-At-Heights Construction. AI, 7(11), 449. https://doi.org/10.3390/ai7110449
Vancouver
Ho C, Chang T, Guo X. A Localized RAG-LLM System for Safety Compliance Question–Answering in Working-At-Heights Construction. AI. 2026;7(11):449. doi:10.3390/ai7110449
BibTeX
@article{Ho2026,
author = {Ho, Chun-Ling and Chang, Tung-Chiung and Guo, Xingyou},
title = {A Localized RAG-LLM System for Safety Compliance Question–Answering in Working-At-Heights Construction},
journal = {AI},
year = {2026},
volume = {7},
number = {11},
pages = {449},
doi = {10.3390/ai7110449}
}
EUVIMED