AtomicRAG: Atom-Entity Graphs for Retrieval-Augmented Generation
Yanning Hou, Jian Huang, Xiaoshu Chen, Xinwang Liu, Siwei Wang, Dahua Yuan, Sihang Zhou, Ke Liang
arXiv (Cornell University)
Problems Identified (4)
Coarse Chunk Knowledge Representation: Using text chunks as the basic knowledge unit rigidly groups multiple atomic facts and limits flexible retrieval across diverse scenarios.
Relation Extraction Error Sensitivity: Triple-based entity linking is sensitive to relation-extraction errors, causing missing or incorrect reasoning paths and reduced retrieval accuracy.
Coarse Chunk Knowledge Representation: Using text chunks as the basic knowledge unit rigidly groups multiple atomic facts and limits flexible retrieval across diverse scenarios.
Relation Extraction Error Sensitivity: Triple-based entity linking is sensitive to relation-extraction errors, causing missing or incorrect reasoning paths and reduced retrieval accuracy.
Proposed Solutions (5)
Atom-Entity Graph: The paper proposes an Atom-Entity Graph architecture for knowledge representation and indexing based on individual self-contained factual knowledge atoms rather than coarse text chunks.
Existence-Based Entity Edges: The approach uses entity edges that indicate whether a relationship exists instead of relying on extracted relation triples.
PageRank Relevance Filtering: The approach combines personalized PageRank with relevance-based filtering to maintain accurate entity connections and improve reasoning reliability.
AtomicRAG Algorithm: The paper presents AtomicRAG as a retrieval-augmented generation algorithm using Atom-Entity Graphs.
Atom-Entity Graph: The paper proposes an Atom-Entity Graph architecture for knowledge representation and indexing based on individual self-contained factual knowledge atoms rather than coarse text chunks.
Results (3)
Improved Retrieval Accuracy:
Improved Reasoning Robustness:
Theoretical and Benchmark Validation:
Research Domain
Retrieval-Augmented Generation / GraphRAG