Large language models (LLMs) demonstrate exceptional performance in general natural language processing tasks but suffer from hallucination in domain-specific scenarios due to reliance on public internet data rather than expert knowledge. To overcome these challenges, we propose a hybrid question-answering (QA) architecture incorporating domain knowledge graphs. The system performs intent recognition, converts natural language queries to structured format, retrieves information from device fault knowledge graphs, and uses LLMs to generate accurate responses. Experimental results show substantial reduction in hallucination rates and improved accuracy compared to pure LLM approaches, providing a viable path for trustworthy domain-oriented QA systems.