This paper introduces Self-BioRAG, a framework that improves medical reasoning by integrating retrieval, self-reflection, and domain-specific knowledge. The framework is trained on 84k filtered biomedical instruction sets to generate explanations, retrieve domain-specific documents, and self-reflect on generated responses. It utilizes a domain-specific retriever (MedCPT) and a self-reflection language model to enhance performance in biomedical and clinical domains. The framework outperforms existing methods, achieving a 7.2% absolute improvement over state-of-the-art open-foundation models with 7B parameters on three major medical question-answering benchmarks. It also outperforms RAG by 8% in Rouge-1 score on two long-form question-answering benchmarks. Self-BioRAG demonstrates its effectiveness by finding clues in the question, retrieving relevant documents, and using encoded knowledge to answer questions, similar to how medical experts would. The framework is available at https://github.com/dmis-lab/self-biorag and the data and code are provided for training and evaluation. The paper also discusses the importance of domain-specific components such as retrievers, document corpora, and instruction sets in addressing domain-related instructions. The framework is trained on biomedical instruction sets, including information extraction, question answering, summarization, text classification, relation extraction, and multi-choice questions. The framework is evaluated on five open-domain question-answering datasets, including multi-choice and long-form question-answering benchmarks. The results show that Self-BioRAG achieves significant performance gains by leveraging domain-specific knowledge and self-reflection. The paper also discusses the effectiveness of adaptive retrieval in improving performance on biomedical and clinical domains. The framework is designed to conditionally generate text without additional training, balancing the trade-off between multiple preferences. The paper concludes that Self-BioRAG is a promising approach for improving medical reasoning in biomedical and clinical domains.This paper introduces Self-BioRAG, a framework that improves medical reasoning by integrating retrieval, self-reflection, and domain-specific knowledge. The framework is trained on 84k filtered biomedical instruction sets to generate explanations, retrieve domain-specific documents, and self-reflect on generated responses. It utilizes a domain-specific retriever (MedCPT) and a self-reflection language model to enhance performance in biomedical and clinical domains. The framework outperforms existing methods, achieving a 7.2% absolute improvement over state-of-the-art open-foundation models with 7B parameters on three major medical question-answering benchmarks. It also outperforms RAG by 8% in Rouge-1 score on two long-form question-answering benchmarks. Self-BioRAG demonstrates its effectiveness by finding clues in the question, retrieving relevant documents, and using encoded knowledge to answer questions, similar to how medical experts would. The framework is available at https://github.com/dmis-lab/self-biorag and the data and code are provided for training and evaluation. The paper also discusses the importance of domain-specific components such as retrievers, document corpora, and instruction sets in addressing domain-related instructions. The framework is trained on biomedical instruction sets, including information extraction, question answering, summarization, text classification, relation extraction, and multi-choice questions. The framework is evaluated on five open-domain question-answering datasets, including multi-choice and long-form question-answering benchmarks. The results show that Self-BioRAG achieves significant performance gains by leveraging domain-specific knowledge and self-reflection. The paper also discusses the effectiveness of adaptive retrieval in improving performance on biomedical and clinical domains. The framework is designed to conditionally generate text without additional training, balancing the trade-off between multiple preferences. The paper concludes that Self-BioRAG is a promising approach for improving medical reasoning in biomedical and clinical domains.