An Agentic GraphRAG Framework with Causal Evidence Reasoning for Explainable Scientific Information Retrieval and Research Recommendation

Abstract

The number of scientific publications continues to increase, and keeping track of them is getting difficult. There has been an explosion of papers published since the beginning of the pandemic, and using the topically similar but causally irrelevant papers from the conventional RAG pipeline to explain why a specific result is recommended is becoming increasingly hard to find. In contrast, previous GraphRAG approaches simply retrieve a single query and retrieve the answer, whereas the proposed agents form and execute sub-queries, traverse entity-relation sub-graphs, and explicitly check for potential causal connections between concepts, methods, and reported results before generating an answer. Experiments on a held-out evaluation split show that the proposed framework is able to achieve more precise retrieval and faithful explanation results than sparse, dense, and non-agentic GraphRAG baselines, and the evidence chains produced by the proposed framework are interpretable and provide evidence that makes the basis of each recommendation auditable. The framework and evaluation protocol are meant to be used as a reusable basis for scientific information retrieval that can be explained.