EviGraphRAG: Constraint-Verified Retrieval-Augmented Question Answering over Event Knowledge Graphs

Abstract

: Event knowledge graphs support event-centric question answering by linking events to temporal, location, participant, and source-record information, but semantic relevance alone does not guarantee that a selected event satisfies every represented condition. We propose EviGraphRAG, a ranker-agnostic reliability layer that separates candidate ranking from exact verification and explicit abstention; EviGraphRAG-Struct is the default configuration. On the controlled oracle-normalized 4800-question benchmark, it preserves the unverified ranker’s positive-task outputs while attaining 0.9792 abstention accuracy. A separate 600-question set from machine-assisted drafting with deterministic construction and semantic-consistency checks shows realistic extraction noise: parser-based Macro Score is 0.7317 vs. 0.8717 with oracle constraints, with location extraction the largest weakness. After quoted-title-aware actor and location parsing repairs, the fixed predicted-cluster evaluation reaches 0.8448 Macro Score without positive false rejection. These results support a scoped claim: exact conjunction over represented event constraints and linked source-record connectivity can improve selective reliability, while parser quality and candidate retrieval remain limiting factors.