AI-powered application of Retrieval-Augmented Generation (RAG) to support non-expert operators in industrial environments

The manufacturing industry is undergoing a transformative change towards digitalization and advanced technologies. The management, organization and reuse of large, heterogenous amount of data generated during the whole manufacturing process remain key challenges. In parallel, human expertise is rarely converted into structured data that can be exploited for process improvement and daily troubleshooting. Although recent advances in Artificial Intelligence (AI) are promising, general-purpose language models exhibit well-known limitations when operating in regulated industrial domains. Responses may be hallucinated and not aligned with company procedures and standards, resulting in distorted information and struggling decisions-making. Retrieval-Augmented Generation (RAG) paradigm is based on a combination of information retrieval system and Large Language Models (LLMs) generation, coupling two fundamental approaches to ensure traceability into specific knowledge domain documentation. In this thesis, a self-hosted Retrieval-Augmented Generation (RAG) system is tailored to better respond to this specific field of application and is tested at Mind4Lab at the Politecnico di Torino within an Industry 4.0 Yaskawa robotic cell configuration. Pipeline improvements are implemented and a specific pragmatic use case focused on pick-and-place tasks is defined. Software improvements include domain-specific semantic chunking tuned to technical manuals, upgraded embedding models, hybrid retrieval with reranking, explicit source-citation traceability, and latency robustness refinements. The aim of the research is to evaluate the effect and the extent of improvement for novice, non-expert operators in robot programming in an industrial environment. Performance is assessed against a baseline relying solely on technical manuals or supervisor assistance, reflecting common practice. A set of metrics is defined to analyze both the tool's effectiveness and human-centered outcomes of this innovative virtual assistant.