Manually resolving drug-gene relationships from oncology clinical practice guidelines is time-consuming and struggles to keep pace with the rapidly growing volume of cancer-specific recommendations. To address this challenge, the Guideline Graph Extracting Framework (GGEF) is proposed, a rigorously evaluated natural language processing (NLP) framework to automatically extract structured, guideline-specific knowledge from unstructured Chinese oncology guidelines. GGEF integrates three core modules: document preprocessing, guideline-aware large language model (LLM) prompting, and systematic accuracy assessment, thus advancing the state of the art in clinical text mining. The core method presented in this paper, GGEF, was validated on a large-scale evaluation set comprising 350 authoritative guidelines and containing 361 expert-verified gene–drug relation pairs. Experimental results show that GGEF significantly outperforms the end-to-end baseline methods, achieving a precision of 97.26%, recall of 95.95%, hallucination rate of 2.74%, and an F1 score of 0.9660. In the future, a public web interface will be provided to ensure system accessibility, making it a practical resource for the biomedical informatics community.