Abstract
Retrieval-Augmented Generation (RAG) has gained significant popularity due to its ability to ground generative models in specific knowledge sources. It can be especially beneficial in high-speed, safety-critical manufacturing settings, given its capacity to reduce information retrieval time, enhance response accuracy, and minimize the risk of hallucinated answers. Given the recency of RAG as a technological innovation, practical implementations are especially valuable, as they offer evidence of its adaptability to specific domains. In this paper, we investigate the real-world application of a RAG chatbot within an industrial context. Using procedural and quality-related documentation from a fast-moving consumer goods (FMCG) manufacturing plant, supported by expert-annotated reference answers, we experiment with various model configurations. Our deployed pipeline achieves faithfulness of 92.63%, answer relevancy of 91.31%, context precision of 75.20%, and context recall of 86.60%. The configuration and codebase for the proposed framework is available on Github.
Showing the abstract — retrieve the full paper via the Exa API.