The growing complexity of steel manufacturing, loss of veteran operators, and demand for carbon-neutral operation make tacit operational knowledge increasingly difficult to sustain. Standalone large language models (LLMs) cannot adequately address terminology mismatch, schema fragmentation, undocumented tacit knowledge, and non-traceable hallucinations in on-site data. This study proposes a domain knowledge graph-based conversational operation assistant integrating a steel-process ontology with expert–LLM dialogue. An eight-class ontology is introduced, featuring the UnfilledSlot class to explicitly represent missing knowledge. A five-phase methodology constructs an initial graph, automatically fills available slots, prioritizes remaining deficiencies, and generates targeted multiple-choice questions for experts. The completed graph supports NL2SQL, analysis, explainable AI, and response generation, while operational logs reveal new knowledge gaps. Applied to a steel continuous casting process with 8,441 slots, automatic filling achieved 62% (5,233 slots), while 600 bundled expert questions resolved 2,702 additional slots, reducing the unfilled ratio from 38% to 6%. Comparisons with a standalone LLM, retrieval-augmented generation, and an ablation without UnfilledSlot, together with expert evaluations across three graph-completeness levels, confirm that explicit deficiency handling improves response reproducibility and traceability.

This paper examines the role of generative AI and large language models (LLMs) in advancing intelligent manufacturing as we transition from Industry 4.0 to Industry 5.0. We begin by analyzing the current limitations of rule-based and manufacturing data systems in facilitating flexible, human-centric production. Next, we categorize LLM utilization strategies into three methodological axes: fine-tuning domain-specific models, employing general-purpose models through prompt engineering, and utilizing retrieval-augmented generation (RAG), which includes multimodal RAG that integrates sensor and text data. For each strategy, we present representative case studies across key application areas such as asset management, maintenance intelligence, quality control, process optimization, and knowledge- and document-centric support systems. Concurrently, we explore how information modeling and ontology-based knowledge graphs can be integrated with LLMs to enhance structured manufacturing semantics, improve source traceability, and minimize hallucinations. Finally, we summarize the advantages and limitations of each approach and propose future research directions for human-centric manufacturing, including the development of trustworthy LLM pipelines, standardized data schemas, and closer integration between digital twins and LLM-based decision support systems.