Beyond Retrieval: Structuring Evolving and Inconsistent External Knowledge with Proposition Relations for RAG
Abstract
AbstractReal-world knowledge bases evolve over time and accumulate outdated statements, contradictions, and misinformation. LLMs have been reported to be vulnerable to such inconsistent contexts. In this paper, we propose Proposition-Structured RAG (ProStruct-RAG), a framework that continuously structures evolving and inconsistent external knowledge bases using timestamped propositions and inter-proposition relations. ProStruct-RAG represents a document collection as a graph whose nodes are propositions and whose edges capture temporal updates, contradictions, and support relations, retrieves query-relevant knowledge as a relational structure, and uses it for answer generation in RAG. Experiments on multiple open-book temporal QA benchmarks show that ProStruct-RAG outperforms baselines that do not use inter-proposition relations, suggesting that RAG over evolving and inconsistent knowledge benefits from organizing external knowledge before retrieval, rather than relying solely on inference-time reasoning. We release our code and Asteria, a new fictional temporal QA dataset designed to mitigate data contamination.