Retrieval-augmented generation enhanced LLM for industrial anomaly detection

Abstract Industrial anomaly detection serves as a crucial safeguard for maintaining production continuity and ensuring operational safety. This paper proposes a novel framework that synergizes Large Language Models (LLMs) with Retrieval-Augmented Generation (RAG) to enhance current detection capabilities. This architecture specifically targets three persistent impediments involving data scarcity, conflicting operational standards, and the isolation of expert knowledge. To systematically address these challenges, the study implements three specific methodological innovations. Firstly, to overcome data constraints, the framework utilizes a frozen reasoning core, coupled with external retrieval, to identify rare anomalies without relying on extensive labelled datasets. Secondly, to mitigate conflicting operational standards, a context-aware partitioning mechanism segregates divergent regulatory protocols, ensuring diagnostic consistency across varying industrial conditions. Thirdly, to bridge the knowledge integration gap, a Human-in-the-Loop (HITL) protocol transforms interactive expert feedback into computable vectors. Unlike traditional fine-tuning approaches, this architecture integrates the frozen reasoning core with a dynamic knowledge base derived from unstructured authoritative standards. A novel context-partitioning mechanism isolates conflicting operational standards to resolve inconsistencies. Simultaneously, the update protocol converts expert feedback into vector representations for real-time model evolution. Experimental validations on the MVTec AD dataset and AWS Welding Handbook demonstrate the framework’s robustness, achieving 80.6% accuracy in visual anomaly detection and 83.67% correctness in textual diagnostics with 100% reference precision. Comparative evaluations indicate that the proposed method significantly outperforms SOTA general-purpose LLMs, including GPT-5 and Deepseek V3.2exp in zero-shot domain adaptability.