221 - ArchRAG: Attributed Community-based Hierarchical Retrieval-Augmented Generation
Association for Artificial Intelligence 2026, Yingli Zhou, Xilin Liu, Mayuchi Mayuchi, Shu Wang
Underline Science Inc.
Problems Identified (4)
Graph RAG relevance identification: Existing graph-based RAG approaches cannot accurately identify relevant information from the graph for QA tasks.
Graph RAG token cost: Existing graph-based RAG approaches consume many tokens during online retrieval.
Graph RAG relevance identification: Existing graph-based RAG approaches cannot accurately identify relevant information from the graph for QA tasks.
Graph RAG token cost: Existing graph-based RAG approaches consume many tokens during online retrieval.
Proposed Solutions (4)
ArchRAG: ArchRAG is a graph-based RAG approach that augments questions using attributed communities and uses LLM-based hierarchical clustering.
Hierarchical attributed-community retrieval: The method builds a hierarchical index structure for attributed communities and uses an online retrieval method to retrieve relevant graph information for a question.
ArchRAG: ArchRAG is a graph-based RAG approach that augments questions using attributed communities and uses LLM-based hierarchical clustering.
Hierarchical attributed-community retrieval: The method builds a hierarchical index structure for attributed communities and uses an online retrieval method to retrieve relevant graph information for a question.
Results (2)
Improved accuracy and token cost:
Improved accuracy and token cost:
Research Domain
Retrieval-Augmented Generation for question answering over graph data