Hybrid Adaptive Retrieval-Augmented Generation for Context Relevance in Biomedical Question Answering

Abstract

Retrieval-Augmented Generation (RAG) has emerged as an effective paradigm for improving the factual accuracy of large language models by grounding responses in external knowledge. This paper proposes a Hybrid Adaptive RAG framework that combines sparse (BM25) and dense retrieval techniques to enhance context relevance in biomedical question answering using the PubMedQA. The proposed approach dynamically integrates lexical and semantic retrieval strategies to improve document ranking and retrieval quality. Experimental results demonstrate that dense retrieval achieves the highest performance, while the hybrid model significantly outperforms BM25. Although Exact Match scores remain low due to the abstractive nature of answers, the system produces contextually grounded responses with high faithfulness. The findings highlight the effectiveness of hybrid retrieval in improving RAG performance while identifying the need for advanced adaptive mechanisms and improved generation models.

Introduction

It explains that keyword-based search methods often fail to capture meaning, while LLMs, although powerful, can produce hallucinations and lack up-to-date domain knowledge. RAG addresses this by combining retrieval of external documents with LLM-based generation, improving accuracy and contextual relevance. However, current RAG systems depend heavily on either sparse retrieval (BM25, keyword-based) or dense retrieval (semantic embeddings), each with limitations when used alone.

To solve this, the paper proposes a Hybrid Adaptive Retrieval-Augmented Generation (HARAG) framework, which dynamically combines both sparse and dense retrieval using an adaptive weighting mechanism. This allows the system to balance exact keyword matching with semantic understanding based on query characteristics.

The framework includes six components: data preparation, sparse retrieval, dense retrieval, hybrid fusion, answer generation using an LLM (e.g., Flan-T5), and evaluation. Biomedical data (such as PubMedQA) is preprocessed and split into document chunks. BM25 handles lexical matching, while transformer embeddings handle semantic similarity. Their outputs are combined using a query-dependent weighting strategy.

The system then feeds top retrieved documents into an LLM to generate answers with reduced hallucination risk. Performance is evaluated using standard retrieval metrics (Recall@k, MRR, nDCG) and generation accuracy (Exact Match).

The literature review shows that while RAG has improved biomedical and enterprise AI systems, most existing methods still rely on static or heuristic fusion of retrieval methods and lack adaptability. It also highlights gaps in evaluation methods and the need for better integration of sparse and dense retrieval.

Conclusion

In this paper, a Hybrid Adaptive Retrieval-Augmented Generation (RAG) framework is proposed to progress context significance in biomedical question answering using the PubMedQA. Experimental results depict that dense retrieval attained the superlative performance, while the proposed hybrid method enhanced over BM25. Though Exact Match scores endured very low due to the abstractive nature of answers, the system curate contextually grounded responses with high faithfulness. Overall, the proposed work illustrates that hybrid retrieval augments performance, but further optimisation is also required to outperform dense retrieval and enhance answer completeness. Future research will focus on adaptive learning methods for regulating weights and cross-encoder reordering to enhance the efficacy of curating the evidence. Moreover, more powerful LLMs are incorporated with domain-specific fine-tuning to enrich the accuracy and completeness of the curated answers.

References

[1] P. Zhao, H. Liu, Y. Chen, Z. Wang, and X. Zhang, “Retrieval-Augmented Generation for AI-Generated Content: A Survey,” Data Science and Engineering, 2026. [2] C. Sharma, “Retrieval-Augmented Generation: A Comprehensive Survey of Architectures, Enhancements, and Robustness Frontiers,” arXiv preprint arXiv:2506.00054, 2025. [3] W. Su, X. Han, Z. Lin, P. Yu, and Z. Sun, “DRAGIN: Dynamic Retrieval Augmented Generation based on the Real-time Information Needs of Large Language Models,” in Proc. ACL, 2024. [4] Y. Yu, H. Wang, X. Liu, and J. Li, “Evaluation of Retrieval-Augmented Generation: A Survey,” in Proc. ICCBD, 2024. [5] S. Wang, R. Zhang, L. Liu, and K. Chen, “FeB4RAG: Evaluating Federated Search in Retrieval-Augmented Generation,” in Proc. SIGIR, 2024. [6] Y. Wang, Z. Li, H. Zhang, and Q. Liu, “Topology-aware Retrieval Augmentation for Text Generation,” in Proc. CIKM, 2024. [7] Z. Wei, Y. Chen, H. Liu, and X. Zhang, “GARAG: Adaptive Question Answering using Retrieval-Augmented Generation,” in Proc. ICCBD, 2024. [8] T. Zhang, X. Liu, J. Wang, and Y. Li, “RAFT: Adapting Language Models to Domain-Specific Retrieval-Augmented Generation,” arXiv preprint arXiv:2403.10131, 2024. [9] S. Es, J. James, L. Espinosa-Anke, and S. Schockaert, “RAGAS: Automated Evaluation of Retrieval Augmented Generation,” arXiv preprint arXiv:2309.15217, 2023. [10] A. Brown, D. Green, M. Taylor, and S. Wilson, “A Systematic Literature Review of Retrieval-Augmented Generation: Techniques, Metrics, and Challenges,” arXiv, 2025. [11] J. Lála, A. Mallen, S. Asai, and H. Hajishirzi, “PaperQA: Retrieval-Augmented Generative Agent for Scientific Research,” arXiv preprint arXiv:2312.07559, 2023. [12] H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, and others, “LLaMA: Open and Efficient Foundation Language Models,” arXiv preprint arXiv:2302.13971, 2023. [13] J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. Le, and D. Zhou, “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,” in Proc. NeurIPS, 2022. [14] R. Kumar, S. Patel, A. Singh, and V. Sharma, “Retrieval-Augmented Generation and LLMs for Enterprise Knowledge Management,” Applied Sciences, 2026. [15] P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W. Yih, T. Rocktäschel, S. Riedel, and D. Kiela, “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks,” arXiv preprint arXiv:2005.11401. [16] Q. Jin, B. Dhingra, Z. Liu, W. W. Cohen, and X. Lu, “PubMedQA: A Dataset for Biomedical Research Question Answering,” in Proc. EMNLP-IJCNLP, 2019, pp. 2567–2577.

Copyright

Copyright © 2026 N Devi, P LeelaRani, V Vidhya, D Jayanthi, Ar. GuruGokul. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.