Integrating a Knowledge-Grounded RAG Chatbot for Multi-Tenant Industrial Machinery Troubleshooting

Abstract

Industrial organizations depend heavily on machinery for daily operations, where even minor failures can cause significant downtime and financial loss. A central challenge in troubleshooting is that the necessary knowledge is dispersed across lengthy manuals and technical documents, making it slow and inefficient for technicians to locate relevant instructions during urgent situations. While recent progress in large language models (LLMs) has demonstrated strong capabilities in understanding and generating natural language, their direct use in industrial environments remains limited. In addition, LLMs lack access to proprietary machine documentation and may produce inaccurate or overly generalized responses, which creates risks in safety-critical contexts. This thesis develops and investigates an intelligent chat-bot powered by Retrieval Augmented Generation (RAG) for end-to-end machinery troubleshooting, conducted in collaboration with Fatman Oy. The proposed system integrates document retrieval with generative capabilities to provide machine-specific, context-aware responses grounded in technical manuals and operational documentation. A key focus of the work is the design of a multi-tenant retrieval architecture that prevents cross-contamination of information between documents belonging to different customers or machines by applying metadata-based filtering before the retrieval stage. In addition, the system incorporates a structured feedback integration mechanism that allows domain experts to refine the knowledge base, ensuring that outdated manual content can be corrected and that the most up-to-date troubleshooting guidance is returned to users. The study also investigates how such a RAG-based chatbot can be deployed using a scalable architecture and seamlessly integrated into an existing application at Fatman Oy, addressing challenges related to document ingestion, vector storage, model hosting, user session management, and workflow coordination. By grounding generative AI in continuously refined documentation within a scalable environment, the proposed approach demonstrates how industrial troubleshooting can be accelerated, information reliability improved, and operational downtime reduced.