The injection molding industry increasingly depends on tacit, experience-driven knowledge, yet this knowledge is becoming difficult to sustain due to workforce aging and the growing presence of multilingual operators. At the same time, conventional AI-based decision-support systems often rely on static documentation or narrowly trained models, limiting their adaptability in real manufacturing environments. This work presents a large language model (LLM)-driven multi-agent framework designed to support practical knowledge transfer and decision-making in injection molding processes. In addition, a domain-specific benchmark is proposed to systematically evaluate expert-system capabilities under realistic manufacturing scenarios. The proposed system combines sparse documented resources, such as troubleshooting tables and manufacturing manuals, with a data-driven process condition generator that infers suitable molding parameters from environmental inputs including temperature and humidity. By orchestrating retrieval-augmented generation (RAG) and tool-calling agents within a modular workflow, the system enables flexible task execution without requiring additional fine-tuning. The framework was evaluated and iteratively refined using a custom-designed benchmark tailored to practical injection molding scenarios, encompassing both standalone and multi-step tasks that require reasoning across multiple information sources and tools. The evaluation results indicate that the multi-agent architecture effectively supports complex, tool-integrated reasoning and demonstrates robust performance across diverse task settings. Moreover, the use of task-oriented, domain-specific benchmarks is essential for moving beyond ad hoc demonstrations toward reproducible, extensible, and industrially meaningful evaluation frameworks, thereby supporting the long-term evolution of AI-assisted expert systems in manufacturing domains.