LLM-Driven Lean Manufacturing System: Integrating AI, RAG, and Embeddings for Industry 4.0 Decision Support

Abstract

The integration of Large Language Models (LLMs) with Lean Manufacturing methodologies represents a significant frontier in the evolution of intelligent manufacturing systems within the Industry 4.0 paradigm. This comprehensive research report presents the architecture, design rationale, and empirical evaluation of a web-based Lean Manufacturing System (LMS) that employs LLMs to automate the generation of structured Lean tools, including Value Stream Mapping (VSM), 5W2H action planning, Ishikawa diagrams, Pareto analysis, GUT prioritization matrices, PDCA cycles, Kanban boards, process flowcharts, and Five Whys root-cause analysis, from unstructured natural-language problem descriptions provided by industrial operators. The proposed architecture adopts a modular microservices design, a Retrieval-Augmented Generation (RAG) pipeline, built upon LangChain, FAISS vector stores, and sentence-transformer embeddings, grounding model outputs in curated domain knowledge sourced from internal project documentation and Lean Manufacturing references. Structured prompt engineering with explicit persona definitions, mathematical output constraints, and JSON schema enforcement constitutes the primary mechanism for deterministic output generation. Experimental validation using a representative industrial case, a connector assembly workstation exhibiting a 12% rework rate, demonstrates that the AI-generated Lean artifacts correctly identify root causes, prioritize countermeasures, and project a 65% lead time reduction alongside a 200% productivity gain. Critical technical limitations are identified and discussed, encompassing single-scenario generation, absence of a feedback loop for continuous learning, dependency on prompt-level logic for semantic consistency, and a lack of integration with real-time shop-floor data streams. We conclude the analysis with a roadmap for overcoming these limitations through multi-agent orchestration, reinforcement learning from human feedback, and dynamic sensor integration.