In policy research, massive unstructured policy texts contain rich value. Nevertheless, their complex semantics and decentralized characteristics make exploring them fully through traditional analysis methods difficult. In this study, we innovatively propose a policy intelligence analysis framework based on a large language model (LLM) and knowledge graph and optimize the model training process by introducing the prompt mechanism (QKV vector bootstrapping), which significantly improves the accuracy of semantic understanding and effectively reduces model illusions. First, the domain adaptive method is adopted to clean the policy text, which is combined with BERT model pretraining to generate standard semantic features; subsequently, the “policy-agency-topic” triad knowledge graph is constructed to realize the visual association reasoning of policy elements. The system adopts a layered architecture design, the front-end is based on React to realize an interactive retrieval interface, and the back-end integrates the Neo4j graph database and the Qwen-Max big model to support intelligent Q&A and policy association analysis. Experiments show that the method significantly improves the accuracy of policy element extraction and the efficiency of the Q&A response. The research results provide an efficient analysis tool for policymakers and researchers, which helps in understanding the connotations and correlations of policies in detail.