Tumor coding classification (TCC) aims to systematically classify and code tumor diseases based on their characteristics, facilitating the statistical analysis and communication of medical information. Existing research typically relies on training deep models to learn normal distributions, whether in supervised or unsupervised settings. However, these training-based methods are often domain-specific, leading to high deployment costs, as any changes in the domain require new data collection and model retraining. In this paper, we radically depart from previous approaches and propose a tumor coding classification method based on large language models (LLMs), introducing a novel, training-free paradigm called TCC-LLMs. This method integrates a re-trainedretrieval-augmented generation pipeline with LLMs to enhance the accuracy of malignant tumor coding classification tasks. Specifically, we first use pre-trained LLMs to extract key pathological staging features from patient records. Then, based on these features and a pre-trained vector database, we retrieve tumor codes that match the pathological stages. Finally, we leverage LLMs to reorder the retrieved tumor codes and determine the final code. Our method achieved an accuracy of 89.26% in the 2024 China Health Information Processing Conference evaluation task, securing fourth place. Our code is available at https://github.com/MagicCat-AI/Chip2024-Tumor-Coding-Classification .