LLM-augmented multi-agent cooperative framework for medical case retrieval in cardiology

Abstract

Retrieving relevant medical cases or documents is a critical information retrieval (IR) task in clinical decision support, particularly in cardiology, yet traditional search methods struggle with complex semantic queries in healthcare. Recent advances in large language models offer powerful language understanding, but LLMs alone cannot reliably retrieve factual cases due to knowledge cutoffs and hallucinations. We focus on a case retrieval task – given a textual description of a patient case, find similar prior cases or pertinent literature – formulated as a general IR problem rather than a purely medical study. Conventional lexical methods often miss semantic similarities, while static dense retrievers falter on out-of-domain medical vocabularies. LLMs can comprehend queries and context, but without external knowledge they may produce inaccurate or non-transparent results. We propose a novel LLM-augmented multi-agent retrieval framework that marries an LLM with dedicated retrieval agents in an iterative cooperation mechanism. Our method uses a LLM as a “planner” agent to reformulate queries and integrate medical (e.g., cardiology) context, and a retrieval agent (with a knowledge index) to fetch candidate cases; the agents interact iteratively, refining search and reranking results via a retrieval-augmented generation (RAG) loop. This multi-agent design contributes three innovations: (1) an iterative query refinement strategy guided by LLM reasoning chains; (2) a cooperative retrieval architecture where an LLM agent and a search agent exchange information to improve relevance; (3) an LLM-based relevance estimator that grounds the LLM with retrieved evidence to mitigate hallucinations. Experiments on three open medical text datasets show our method outperforms baseline models by 5.3–6.1 percentage points in Recall@10 and NDCG, with statistically significant gains. We also observe improved generalization to novel conditions and robustness to query noise compared to baselines. The proposed framework, while validated on medical text, is broadly applicable to other knowledge-intensive retrieval tasks (legal case search, technical support archives), providing a foundation for intelligent IR systems that leverage both learning-based understanding and explicit retrieval for transparency and up-to-date knowledge.