A graph-based retrieval method is introduced to address the challenge of capturing long-tail biomedical knowledge, which is often overlooked by large language models (LLMs) due to their tendency to focus on frequently seen information. The study proposes a novel knowledge graph-based retrieval approach that mitigates the information overload problem by downsampled clusters of overrepresented concepts in biomedical literature. This method outperforms traditional embedding similarity-based retrieval in terms of precision and recall. The approach leverages a knowledge graph to structure biomedical entities and relationships, enabling more effective retrieval of rare and recent discoveries. The study also demonstrates that combining embedding similarity and knowledge graph retrieval methods in a hybrid model leads to improved performance in biomedical question-answering. The results show that the knowledge graph-based retrieval method significantly outperforms embedding similarity-based retrieval, particularly in retrieving relevant information from the long tail of biomedical knowledge. The study highlights the importance of data balancing in retrieval systems and suggests that integrating knowledge graphs with LLMs can enhance the accuracy and comprehensiveness of information retrieval in biomedical research.A graph-based retrieval method is introduced to address the challenge of capturing long-tail biomedical knowledge, which is often overlooked by large language models (LLMs) due to their tendency to focus on frequently seen information. The study proposes a novel knowledge graph-based retrieval approach that mitigates the information overload problem by downsampled clusters of overrepresented concepts in biomedical literature. This method outperforms traditional embedding similarity-based retrieval in terms of precision and recall. The approach leverages a knowledge graph to structure biomedical entities and relationships, enabling more effective retrieval of rare and recent discoveries. The study also demonstrates that combining embedding similarity and knowledge graph retrieval methods in a hybrid model leads to improved performance in biomedical question-answering. The results show that the knowledge graph-based retrieval method significantly outperforms embedding similarity-based retrieval, particularly in retrieving relevant information from the long tail of biomedical knowledge. The study highlights the importance of data balancing in retrieval systems and suggests that integrating knowledge graphs with LLMs can enhance the accuracy and comprehensiveness of information retrieval in biomedical research.