Event-state knowledge graph-enhanced GraphRAG for order-driven 3D printing supply chain configuration

In large-scale mass-customized 3D printing supply chains, lead enterprises are required to decompose complex customer orders into multiple manufacturing tasks and execute them collaboratively across distributed factory clusters. However, owing to the lack of unified modeling and reasoning mechanisms for order decomposition processes and cluster-level resources, order-driven multi-factory collaborative supply chain configuration remains a significant challenge. This paper proposes an intelligent supply chain configuration method that integrates an event-state knowledge graph (ESKG) with graph-based retrieval-augmented generation (GraphRAG). The method first constructs a dynamically updatable ESKG by defining six categories of event-state descriptive triples. Subsequently, a BS-LRTE hybrid extraction method is introduced to extract the aforementioned triples. This method employs a BERT-BiLSTM-CRF model as a structural-constraint module and a large language model (LLM) as a semantic-reasoning and completion module, thereby enhancing the recognition of implicit engineering semantics while mitigating hallucination in generative models. Finally, an ESKG-GraphRAG-based hybrid retrieval-augmented mechanism is developed, in which the ESKG serves as GraphRAG’s local knowledge base to enable synergy between textual semantic retrieval and structured reasoning. Ablation studies are conducted to systematically analyze the contributions of key components, including ESKG, GraphRAG, and LLMs, to recommendation performance. In addition, comparative analyses of LLMs with different scales are performed using representative case studies. Experimental results demonstrate that the proposed framework achieves accurate supply chain configuration under both normal and abnormal scenarios.