RAG-ReAct empowered large language models for localized intelligent bridge operation and maintenance: a knowledge-driven agentic approach

Bridges serve as critical transportation infrastructure hubs. The recent breakthroughs in large language models (LLMs) offer transformative potential for intelligent bridge operation and maintenance (O&M), but their direct application is hindered by domain-specific hallucination, black-box reasoning opacity, and passive response limitations. This manuscript proposes a knowledge-driven, edge-side bridge O&M agent based on LLMs, addressing those critical limitations of general-purpose LLMs in specialized structural health monitoring (SHM) domains. The methodology centers on three core technical components: (1) an automated knowledge base construction pipeline; (2) a standardized action space generated via LLMs, wherein each O&M tool is encapsulated with natural language descriptions; and (3) an integrated retrieval-augmented generation (RAG)–reasoning acting (ReAct) agentic workflow that executes closed-loop O&M tasks through “knowledge-driven thought–action execution–observation feedback” cycles. Systematic experiments on the Qwen3 series (0.6B–14B parameters) with a 60,000-character highway bridge O&M knowledge base demonstrate that domain knowledge base integration significantly enhances professional accuracy, and the proposed agent achieves effective closed-loop bridge O&M workflow execution on a single edge-side device, offering a lightweight, specialized, and proactive technical solution for intelligent bridge O&M.