A Multi-Stage Ranking Pipeline for High-Precision Medical Information Retrieval

Effective retrieval of biomedical information presents a significant challenge due to terminological complexity and semantic ambiguity. Traditional keyword-based methods like BM۲۵ often fail to capture the user's semantic intent. To address this, we propose and empirically evaluate a multi-stage ranking architecture designed for high- precision retrieval. Our pipeline initiates with two parallel retrieval stages: a sparse lexical retriever (BM۲۵) and a dense semantic retriever using a Bi-Encoder model (multi-qa-MiniLM-L۶-cos-v۱). The resulting candidate lists are then fused using Reciprocal Rank Fusion (RRF) to leverage their complementary strengths. In the final stage, a more powerful Cross-Encoder model (ms-marco-MiniLM-L- ۶-v۲) re-ranks the top ۱۰۰ candidates from the fused list to achieve fine-grained relevance scoring. Evaluated on the standard TREC- COVID dataset, our complete pipeline demonstrates substantial performance gains at each stage, culminating in a final Precision@۱۰ of ۰.۸۰۸ and an nDCG@۱۰ of ۰.۷۵۴. This represents a significant relative improvement of ۶۸% and ۶۹%, respectively, over the BM۲۵ baseline. These results validate the efficacy of a cascaded retrieve- fuse-rerank architecture. Our work underscores the synergistic value of combining sparse, dense, and cross-attention models, providing a robust framework for developing high-performance information retrieval systems in specialized domains.

Keywords:

Information Retrieval , Hybrid Search , Re-ranking , Large Language Models (LLMs)