Overview

I developed the conceptual foundation for a Retrieval-Augmented Generation (RAG) AI agent to assist with machine failure troubleshooting. My goal is to reduce time spent on initial diagnosis by interpreting operator-reported symptoms and delivering relevant, machine-specific guidance sourced from existing owners' manuals and troubleshooting documentation. Rather than relying on rigid decision trees, I designed the system around a symptom-based diagnostic approach that reflects real-world complexity, where similar symptoms can indicate different root causes and a single issue may present in multiple ways depending on context.

Project Objective

My objective is to build a RAG-based AI agent that reduces time spent on initial diagnosis by interpreting operator-reported symptoms and providing relevant, machine-specific troubleshooting guidance. Rather than replacing existing documentation, the system leverages already-developed owners' manuals and troubleshooting guides by converting them into a structured, retrievable knowledge base suitable for AI-driven reasoning. A core design requirement was a symptom-based diagnostic approach that accounts for the fact that the same issue can present with multiple symptoms, and that similar symptoms can indicate different root causes depending on context.

What Was Done

• I developed a canonical symptom taxonomy to normalize free-form operator input, addressing variability in how operators describe issues even when observing the same behavior. The agent is intended to map free-form input to this standardized symptom set before progressing through the diagnostic workflow.

• I designed the system to prompt operators with targeted clarifying questions to refine and confirm observations once symptoms are mapped—for example, distinguishing whether "the machine is not running" refers to a lack of power, a failed startup sequence, or an operational fault, each leading to a different diagnostic path.

• I defined a non-linear, graph-based diagnostic structure that allows multiple symptoms to converge on a single cause and a single symptom to branch into multiple potential causes, avoiding the failure points of rigid, linear decision trees.

• I defined a probabilistic ranking model that combines normalized symptoms with contextual information gathered through clarifying questions (machine state, recent changes, operating conditions) to rank likely causes and guide operators toward the most probable troubleshooting paths first.

End Result

At the current stage, I've established the conceptual framework and diagnostic logic needed to support a RAG-based AI troubleshooting agent—the symptom taxonomy, graph-based diagnostic approach, and probabilistic reasoning model are all defined, and I've identified existing documentation as the primary knowledge source for retrieval. Next steps include parsing and structuring manuals for effective retrieval, implementing the retrieval and ranking mechanisms, and integrating the symptom normalization and clarification workflow into a functional AI agent. Once completed, the system is expected to provide faster, more consistent initial diagnostics by translating subjective operator input into structured reasoning and prioritized troubleshooting guidance.