CWAGraph: Retrieving What Was Never Explicitly Identified in Graph-Based RAG

Unlike chunk-based Retrieval-Augmented Generation (RAG) methods which retrieve information only from text passages, graph-based RAG methods improve their performance through entity-relation graphs. However, existing graph-based RAG methods typically represent graph facts under the Open World Assumption (OWA), where an absent edge (relation) means unknown rather than false. This principle is problematic for many closed corpora including technical manuals and regulations, where rules are generally stated through scope-bounded statements (such as all X, except Y or only Y) named as local completeness statements(LCSs) in this paper. LCS apply rules to every unmentioned in-scope entity which cannot be explicitly and completely represented by the OWA graph. To address this problem, we propose CWAGraph, which is a graph-based RAG framework augmented with a Closed World Assumption (CWA) layer representing local scope and boundaries for LCSs. The CWA layer identifies which entities are governed by each statement and makes the corresponding closed-world evidence retrievable from those entities, thus enhancing the performance of question answering (QA) involving LCS (called as LCS resolution in this paper). To better evaluate RAG systems' performance on LCS resolution, we further introduce a new QA benchmark SetQA consisting of 493 QA instances involving LCSs across two source domains and three difficulty levels. Our experiments on SetQA reveal that CWAGraph consistently outperforms chunk-based and graph-based RAG baselines.