AtomicRAG: Enhancing Retrieval-Augmented Generation with Atom-Entity Graphs

The recent advancements in GraphRAG methodologies have incorporated graph structures into the processes of text indexing and retrieval. By utilizing knowledge graph triples, these methods aim to enhance the coverage and precision of information retrieval. However, a significant limitation arises when treating text segments as the primary unit of knowledge representation, as this approach tends to cluster multiple atomic facts together. This rigidity restricts the adaptability required for various retrieval contexts. Furthermore, the reliance on triple-based entity linking is prone to errors in relation extraction, which can result in incomplete or inaccurate reasoning pathways, ultimately diminishing retrieval effectiveness. To overcome these challenges, we introduce the Atom-Entity Graph, a refined architecture for knowledge representation and indexing. Our method emphasizes knowledge atoms—distinct, self-sufficient units of factual information—allowing for flexible reconfiguration without interference. Relationships between entities are represented simply as existing connections. By integrating personalized PageRank with relevance-based filtering, we ensure accurate entity relationships and enhance reasoning reliability. Theoretical evaluations and experiments across five public datasets demonstrate that the AtomicRAG algorithm surpasses existing RAG benchmarks in both retrieval accuracy and reasoning strength.