Authors: Pankaj Rajesh Bhakde, Professor Abhishek Popat Bangar, Professor S. B. Bhosale, Professor V. A. Karad
Abstract: IoT deployments bring together a wide mix of sensors, gateways, embedded devices, and software services. Finding the cause of a fault, or setting up a device the technician has never seen before, calls for knowledge that is typically spread across datasheets, manuals, incident tickets, and past reports. This review looks at retrieval-augmented generation (RAG) and large language model (LLM) techniques for IoT troubleshooting, using five selected studies that cover incident diagnosis and reporting, technical-data extraction, agentic answer validation, an industrial troubleshooting assistant, and conversational machine monitoring. Across these works, retrieval anchors generated answers to outside evidence, structured extraction makes technical specifications easier to query, and validation steps cut down on mix-ups between similarly named fields and on answers the evidence does not support. The reported evaluations show measurable improvements in technical question answering and task performance, yet they also expose weaknesses with hard questions, varied document layouts, source reliability, and generalisation to new deployments. We compare the methods and contributions of the studies, point out research gaps, and outline an evidence-centred architecture that joins document ingestion, structured device knowledge, hybrid retrieval, sensor-context reasoning, answer verification, and technician feedback. Our main conclusion is that a dependable troubleshooting chatbot cannot rest on fluent text generation alone: its advice has to be traceable to device-specific evidence and bounded by technical rules. Since the review draws on five chosen papers rather than an exhaustive database search, it is offered as a focused narrative review, not a systematic one.
