Wind turbine maintenance questions require both equipment relationships and detailed operational knowledge. Document retrieval can miss fault dependencies, whereas a knowledge graph may omit inspection procedures and operating conditions. This paper presents a collaborative retrieval workflow that combines a maintenance knowledge graph with document retrieval-augmented generation. Aho–Corasick matching and LLM-assisted Dice alignment identify graph entities; schema-guided Cypher retrieves related faults, causes, and measures. Document retrieval supplements incomplete graph answers before final synthesis. The prototype integrates a graph containing 3,841 nodes and 4,443 relationships with 29 maintenance documents. On 150 domain questions, the collaborative workflow obtained weighted scores of 83.2, 85.4, and 81.3 from DeepSeek, Qwen, and Kimi judges, respectively, exceeding the base LLM and both single-source retrieval workflows. Relative gains over the base LLM ranged from 25.3% to 30.1%. These results support the complementary use of structured and textual knowledge, while module-level contributions and field reliability require further validation.

wind turbine maintenance,retrieval-augmented generation,entity alignment,question answering,knowledge graph