Knowledge-intensive question answering in high-stakes domains such as medicine, law, and finance demands systems that deliver accurate, verifiable, and temporally current information. While large language models (LLMs) exhibit remarkable generative fluency, they are fundamentally constrained by fixed knowledge cutoffs, hallucination, and limited source attribution — rendering them unreliable for decision-critical applications where factual precision, regulatory compliance, and auditability are essential. This paper presents a Knowledge-Adaptive Retrieval-Augmented Generation (KA-RAG) framework implemented on Amazon Bedrock, addressing these limitations by grounding generative responses in dynamically maintained, domain-partitioned retrieval indices rather than static parametric memory. The framework integrates four core components: hybrid retrieval combining Titan Embeddings dense search with OpenSearch BM25; cross-encoder re-ranking for precision-optimised passage selection; dynamic knowledge fusion for multi-document context assembly; and citation-augmented generation with Claude 3 and Llama 3, producing responses with token-level source attribution for full verifiability.
A central design principle of KA-RAG is knowledge adaptivity: incoming queries are automatically classified by domain and routed to independently maintained, versioned knowledge partitions, ensuring retrieval operates over domain-coherent corpora rather than mixed-domain indices. This design respects domain-specific data governance and multi-tenant isolation requirements critical for regulated industries. Bedrock Guardrails are integrated as a factuality verification layer, applying multi-stage consistency checks between retrieved evidence and generated claims to suppress hallucinated content before responses reach end users. The serverless Bedrock infrastructure further eliminates provisioning overhead and reduces knowledge update latency from days to hours through automated ingestion pipelines.
Experiments across three benchmark datasets — MedQA (medical), LegalBench (legal), and FinQA (financial) — demonstrate substantial improvements over competitive RAG baselines: a 34.2% increase in exact match accuracy, a 28.7% improvement in factual consistency, and a 41.3% reduction in hallucination rates. Ablation studies confirm that each architectural component contributes independently to overall gains. The framework’s token-level attribution mechanism additionally enables compliance officers to audit every response against its source evidence — a capability increasingly demanded by regulatory frameworks governing AI in healthcare and financial services. These results establish that cloud-native, retrieval-augmented architectures designed with domain adaptivity and operational transparency as first-class concerns can meet the reliability standards required for enterprise deployment.
Keywords
Retrieval-Augmented Generation, Question Answering, Amazon Bedrock, Knowledge Bases, Hallucination Mitigation, Domain Adaptation, Large Language Models
Conclusion
The Knowledge-Adaptive Retrieval-Augmented Generation (KA-RAG) framework implemented on Amazon Bedrock demonstrates that domain-specific question answering can be both highly accurate and operationally scalable when retrieval, re-ranking, dynamic fusion, and citation-aware generation are tightly integrated within a managed cloud ecosystem. The system consistently outperforms traditional IR-based QA, dense-only retrieval, and prior RAG architectures across medical, legal, and financial domains, achieving a 34.2% improvement in exact match accuracy, 28.7% improvement in factual consistency, and 41.3% reduction in hallucination rates.
The Bedrock-native implementation provides serverless vector search, automated knowledge ingestion, multi-model orchestration, and enterprise-grade governance, reducing operational burden and enabling continuous knowledge updates without retraining. Despite challenges in multi-hop reasoning and domain adaptation costs, this work demonstrates that combining RAG with Amazon Bedrock’s managed AI ecosystem provides a powerful foundation for trustworthy, domain-adaptive QA systems suitable for knowledge-intensive and safety-critical domains.
References
[1] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding,” Proceedings of NAACL-HLT, 2019.
[2] D. Chen, A. Fisch, J. Weston, and A. Bordes, “Reading Wikipedia to Answer Open-Domain Questions,” Proceedings of ACL, 2017.
[3] P. Lewis, E. Perez, A. Piktus, et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks,” NeurIPS, 2020.
[4] T. Brown, B. Mann, N. Ryder, et al., “Language Models Are Few-Shot Learners,” NeurIPS, 2020.
[5] G. Izacard and E. Grave, “Leveraging Passage Retrieval With Generative Models for Open-Domain Question Answering,” ACL, 2021.
[6] S. Borgeaud, A. Mensch, J. Hoffmann, et al., “Improving Language Models by Retrieving From Trillions of Tokens,” ICML, 2022.
[7] V. Karpukhin, B. Oguz, S. Min, et al., “Dense Passage Retrieval for Open-Domain Question Answering,” EMNLP, 2020.
[8] O. Khattab and M. Zaharia, “ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction,” SIGIR, 2020.
[9] L. Xiong, C. Xiong, Y. Li, et al., “Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval,” ICLR, 2021.
[10] T. Gao, X. Yao, and D. Chen, “SimCSE: Simple Contrastive Learning of Sentence Embeddings,” EMNLP, 2021.
[11] S. Gururangan, A. Marasovic, S. Swayamdipta, et al., “Don’t Stop Pretraining: Adapt Language Models to Domains and Tasks,” ACL, 2020.
[12] J. Howard and S. Ruder, “Universal Language Model Fine-Tuning for Text Classification,” ACL, 2018.
[13] V. Sanh, L. Debut, J. Chaumond, and T. Wolf, “DistilBERT: A Distilled Version of BERT,” NeurIPS Workshop, 2019.
[14] W. Kryscinski, B. McCann, C. Xiong, and R. Socher, “Evaluating the Factual Consistency of Abstractive Summarization,” EMNLP, 2020.
[15] K. Shuster, S. Poff, M. Chen, et al., “Retrieval-Augmented Generation for Knowledge-Grounded Dialogue,” EMNLP, 2021.
[16] J. Kuhn, et al., “Uncertainty-Aware Language Models,” arXiv:2302.09664, 2023.
[17] H. Rashkin, et al., “Measuring Attribution in Language Models,” ACL, 2023.
[18] A. Asai, et al., “Self-RAG: Learning to Retrieve, Generate, and Critique Through Self-Reflection,” arXiv:2310.11511, 2023.
[19] G. Izacard, et al., “Atlas: Few-Shot Learning With Retrieval-Augmented Language Models,” arXiv:2208.03299, 2022.
[20] K. Guu, K. Lee, Z. Tung, et al., “REALM: Retrieval-Augmented Language Model Pre-Training,” ICML, 2020.
[21] D. Jin, et al., “What Disease Does This Patient Have? A Large-Scale Open-Domain Medical QA Dataset,” EMNLP, 2021.
[22] N. Holzenberger, et al., “LegalBench: A Benchmark for Legal Reasoning in Large Language Models,” arXiv:2308.11462, 2023.
[23] Z. Chen, et al., “FinQA: Numerical Reasoning Over Financial Data,” EMNLP, 2021.
[24] Amazon Web Services, “Amazon Bedrock Documentation,” AWS, 2023.
[25] Amazon Web Services, “Amazon OpenSearch Serverless Documentation,” AWS, 2023.
[26] T. Zhang, V. Kishore, F. Wu, et al., “BERTScore: Evaluating Text Generation With BERT,” ICLR, 2020.
[27] C.-Y. Lin, “ROUGE: A Package for Automatic Evaluation of Summaries,” ACL Workshop, 2004.
[28] C. Manning, P. Raghavan, and H. Schutze, Introduction to Information Retrieval, Cambridge University Press, 2008.
How to Cite This Paper
Duru Juliet Chinenye, Ogbuagu Chinedu Samuel, Chima Aguocha Obingonye, Hilary Prince-Daniel Chikaodi (2026). Knowledge-Adaptive Question Answering with Retrieval-Augmented Generation on Amazon Bedrock. International Journal of Computer Techniques, 13(2). ISSN: 2394-2231.
