Sparse-Dense Fusion with Reciprocal Rank Fusion for HallucinationMitigated Clinical Knowledge Retrieval
International Journal of Drug Delivery Technology · 29.06.2026 · 16 (59s)
DOI: 10.25258/ijddt.16.59s.94 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Drug Delivery Technology, online erschienen am 29.06.2026. Verfasst von Abhay Nautiyal und 2 weiteren Autor:innen.
Abstract (Originalsprache)
Large language models (LLMs) hallucinate—they generate confident but unverifiable medical information, a critical failure in healthcare where accuracy directly impacts patient outcomes. This paper presents a hybrid RetrievalAugmented Generation (RAG) framework that mitigates hallucination by grounding generated answers in verified medical context. The system integrates sparse BM25 retrieval for precise medical term matching with dense FAISS retrieval for semantic similarity through Reciprocal Rank Fusion (RRF), ensuring that neither exact terminology nor conceptual relatedness is overlooked. A cross-encoder reranker refines the fused results, and Llama 3.1 8B Instant is conditioned on top-ranked documents to produce answers with explicit source attribution. Evaluated on 16,412 questionanswer pairs from the MedQuAD dataset covering 138 medical focus areas, the hybrid approach achieves Precision@10 = 63.80%, Recall@10 = 61.84%, NDCG@10 = 78.07%, and Mean Reciprocal Rank = 94.87%—significantly outperforming individual BM25 and FAISS baselines (p < 0.01, bootstrap test). Analysis reveals that 83% of retrieved items change rank after cross-encoder reranking, with the top-1 document changing in 55% of queries, while method overlap analysis demonstrates complementary retrieval behavior with only 42% agreement across methods. The complete system is deployed as an interactive Gradio web application enabling real-time multi-method comparison for clinical decision support.
Zitieren
APA
Nautiyal, A., Kumar, R., & Thakkar, V. M. (2026). Sparse-Dense Fusion with Reciprocal Rank Fusion for HallucinationMitigated Clinical Knowledge Retrieval. International Journal of Drug Delivery Technology, 16(59s). https://doi.org/10.25258/ijddt.16.59s.94
Vancouver
Nautiyal A, Kumar R, Thakkar VM. Sparse-Dense Fusion with Reciprocal Rank Fusion for HallucinationMitigated Clinical Knowledge Retrieval. International Journal of Drug Delivery Technology. 2026;16(59s). doi:10.25258/ijddt.16.59s.94
BibTeX
@article{Nautiyal2026,
author = {Nautiyal, Abhay and Kumar, Ramesh and Thakkar, V. M.},
title = {Sparse-Dense Fusion with Reciprocal Rank Fusion for HallucinationMitigated Clinical Knowledge Retrieval},
journal = {International Journal of Drug Delivery Technology},
year = {2026},
volume = {16},
number = {59s},
doi = {10.25258/ijddt.16.59s.94}
}
EUVIMED