This study proposes the application of large language models (LLMs) for the automatic annotation of medical texts in Portuguese, with a specific focus on the oncology domain. To address the need for structuring non-standardized information in electronic health records, this research proposes a methodology that integrates deep learning techniques, manual annotation, and preprocessing strategies to improve entity recognition. By employing models such as GLiNER, BioBERT pt, and GPT-4o, the study compares their effectiveness in extracting terms related to diseases, medications, and medical procedures. Notably, GPT-4o demonstrated superior contextual adaptability, particularly when utilizing few-shot learning, enabling precise identification of complex oncological entities. The findings highlight that integrating multiple models can enhance data consistency and quality while reducing both the costs and time associated with manual annotation.