Hybrid-IR: Dual-Path Hybrid Retrieval with Iterative Reasoning for Complex Medical Question Answering

Medical question answering (QA) systems often struggle with two major issues: hallucinations and the inability to synthesize fragmented knowledge across multiple documents. While standard Retrieval-Augmented Generation (RAG) helps, most existing methods rely on a single retrieval path—either vector-based or graph-based—and static retrieval strategies that fail to support the multi-step reasoning required for complex clinical queries.

The authors propose "Hybrid-IR," a framework that addresses these limitations through a dual-path retrieval mechanism combined with an iterative reasoning loop.

1. Dual-Path Indexing: The system creates a hybrid index consisting of a Knowledge Graph (KG) that captures structured entity-relation dependencies and a vector index that preserves fine-grained semantic details from original text.
2. Iterative Retrieval-Reasoning: Instead of a single retrieval step, the model uses a "retrieve-reason-retrieve" loop. It decomposes complex questions into sub-questions, retrieves evidence from both paths, generates sub-answers, and uses the accumulated history to refine subsequent retrieval and reasoning steps.
3. Fusion: The system uses Reciprocal Rank Fusion (RRF) to combine the results from the graph-based and dense retrieval paths, ensuring that both structural and semantic relevance are considered.

Experiments across three medical QA benchmarks demonstrate that Hybrid-IR significantly outperforms standard RAG baselines. By enabling dynamic, multi-step evidence gathering, the framework improves the accuracy of backbone LLMs (such as Llama-3-8B-Instruct and GPT-4o-mini) by up to 10 percentage points. The results suggest that integrating structured relational knowledge with unstructured semantic retrieval is critical for complex biomedical reasoning.