A2RAG: Adaptive Agentic Graph Retrieval for Cost-Aware and Reliable Reasoning
Jiate Liu, Zhengyi Yang, Jingling Wu, Mingchen Ju, Xin Cao, Danting Zhang, Shaobo Qiao, Guanfeng Liu, Bocheng Han, Shuyue Yu, Xin Shu, Zebin Chen, Dong Wen
Open MIND
Problems Identified (4)
Mixed-difficulty retrieval inefficiency: Practical Graph-RAG deployments suffer when one-size-fits-all retrieval wastes cost on easy queries or fails on hard multihop cases.
Graph extraction loss: Graph abstraction can omit fine-grained qualifiers that remain only in source text, reducing reliability under incomplete graph representations.
Mixed-difficulty retrieval inefficiency: Practical Graph-RAG deployments suffer when one-size-fits-all retrieval wastes cost on easy queries or fails on hard multihop cases.
Graph extraction loss: Graph abstraction can omit fine-grained qualifiers that remain only in source text, reducing reliability under incomplete graph representations.
Proposed Solutions (5)
A2RAG adaptive agentic GraphRAG: A2RAG is an adaptive-and-agentic GraphRAG framework designed for cost-aware and reliable reasoning.
Evidence-sufficiency adaptive controller: The framework uses an adaptive controller to verify evidence sufficiency and trigger targeted refinement only when needed.
Agentic provenance-text retriever: The framework uses an agentic retriever that escalates retrieval effort and maps graph signals back to provenance text for robustness under extraction loss and incomplete graphs.
A2RAG adaptive agentic GraphRAG: A2RAG is an adaptive-and-agentic GraphRAG framework designed for cost-aware and reliable reasoning.
Evidence-sufficiency adaptive controller: The framework uses an adaptive controller to verify evidence sufficiency and trigger targeted refinement only when needed.
Results (3)
Higher Recall@2:
Reduced token consumption and latency:
Higher Recall@2:
Research Domain
Graph Retrieval-Augmented Generation for multihop question answering