Scholarlink

The explosive growth of biomedical literature in PubMed creates a challenge for the retrieval of accurate and context-sensitive information. To address this, we introduce a PubMed-integrated Retrieval-Augmented Generation (RAG) framework that integrates Google's Gemini 1.5 Flash with semantic storage in ChromaDB. Biomedical abstracts are dynamically queried using NCBI E-utilities, pre-processed using XML parsing, cleaning, and chunking, prior to being semantically embedded using Google Generative AI embeddings. The resultant embeddings are stored in Chroma DB to facilitate effective similarity-based retrieval and context-aware response generation. Upon receiving user queries, the system enriches them with appropriate PubMed fragments and uses Gemini to generate response with supporting citation. Experimental analysis reports an overall accuracy of 81%, demonstrating a reduction in hallucination on biomedical queries. Conversely, Gemini without retrieval (no RAG) attained 62% accuracy, highlighting the efficacy of the new method. Qualitative case studies, such as papillary thyroid cancer and Paget disease, also reinforce factual grounding, traceability of citations, and openness. The framework thus shows an evidence-based, scalable assistant to facilitate biomedical research, clinical decision support, and higher education.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Scalable Retrieval-Augmented Generation Pipeline for Reliable Biomedical Question Answering

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • S. Kishore Balaji,
  • R. Bhuvanya,
  • R. Mahalakshmi,
  • M. Kaviya

摘要

The explosive growth of biomedical literature in PubMed creates a challenge for the retrieval of accurate and context-sensitive information. To address this, we introduce a PubMed-integrated Retrieval-Augmented Generation (RAG) framework that integrates Google's Gemini 1.5 Flash with semantic storage in ChromaDB. Biomedical abstracts are dynamically queried using NCBI E-utilities, pre-processed using XML parsing, cleaning, and chunking, prior to being semantically embedded using Google Generative AI embeddings. The resultant embeddings are stored in Chroma DB to facilitate effective similarity-based retrieval and context-aware response generation. Upon receiving user queries, the system enriches them with appropriate PubMed fragments and uses Gemini to generate response with supporting citation. Experimental analysis reports an overall accuracy of 81%, demonstrating a reduction in hallucination on biomedical queries. Conversely, Gemini without retrieval (no RAG) attained 62% accuracy, highlighting the efficacy of the new method. Qualitative case studies, such as papillary thyroid cancer and Paget disease, also reinforce factual grounding, traceability of citations, and openness. The framework thus shows an evidence-based, scalable assistant to facilitate biomedical research, clinical decision support, and higher education.