This study addresses the practical needs of UAV equipment training by targeting the inefficiencies of manual annotation in traditional knowledge graph(KG) construction. It researches a large language model-driven automated method for constructing training knowledge graphs. Under the designed graph schema, it proposes a prompting extraction algorithm that integrates Chain-of-Thought (CoT) reasoning and multi-round judgment mechanisms, enabling automated identification and extraction of entities and relations from document-level unstructured training texts. Through comparing extraction performance across different LLM frameworks and conducting ablation experiments, the algorithm's effectiveness is validated. Based on this methodology, a specific UAV model training KGs are constructed; combined with the Chain of Exploration (CoE) retrieval algorithm and interactive training templates, a training assistant agent is developed, verifying the KG’s practical applicability. These research outcomes provide an actionable technical pathway for intelligent equipment training.