Atom-Entity Graphs Improve Retrieval-Augmented Generation
A new paper on arXiv proposes Atom-Entity Graphs, a knowledge representation architecture that stores information as atomic facts rather than text chunks. The authors argue that existing GraphRAG methods, which use knowledge graph triples to connect text chunks, suffer from rigid grouping of multiple facts and sensitivity to relation-extraction errors. Their approach stores knowledge as individual, self-contained units called knowledge atoms, allowing flexible retrieval and more reliable reasoning paths. The paper is available at https://arxiv.org/abs/2604.20844.
Key facts
- Paper proposes Atom-Entity Graphs for retrieval-augmented generation.
- Knowledge is stored as atomic facts rather than text chunks.
- Existing GraphRAG methods use knowledge graph triples to connect text chunks.
- Atom-Entity Graphs aim to improve retrieval flexibility and accuracy.
- The approach addresses issues with rigid grouping and relation-extraction errors.
- Paper available on arXiv with ID 2604.20844.
- Published on arXiv as a cross-type announcement.
- The method allows knowledge elements to be flexibly combined.
Entities
Institutions
- arXiv