Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A2RAG: Adaptive Agentic Graph Retrieval for Cost-Aware and Reliable Reasoning

Jiate Liu, Zebin Chen, Shaobo Qiao, Mingchen Ju, Danting Zhang, Bocheng Han, Shuyue Yu, Xin Shu, Jingling Wu, Dong Wen, Xin Cao, Guanfeng Liu, Zhengyi Yang

TL;DR

A2RAG tackles the challenge of grounding large language models for knowledge-intensive, multi-hop QA by separating reliability control from progressive evidence acquisition. It combines an Adaptive Control Loop that verifies answers against provenance and triggers bounded rewrites with an Agentic Retriever that starts with local graph evidence and escalates to bridge and global diffusion, finally map-back to source text to recover fine-grained qualifiers. The framework delivers strong evidence-retrieval gains (Recall@2) and substantial efficiency improvements, while remaining robust to extraction loss through provenance-based grounding. This approach offers a practical, cost-aware path to reliable graph-augmented reasoning in production settings, where graph abstractions may be imperfect or incomplete.

Abstract

Graph Retrieval-Augmented Generation (Graph-RAG) enhances multihop question answering by organizing corpora into knowledge graphs and routing evidence through relational structure. However, practical deployments face two persistent bottlenecks: (i) mixed-difficulty workloads where one-size-fits-all retrieval either wastes cost on easy queries or fails on hard multihop cases, and (ii) extraction loss, where graph abstraction omits fine-grained qualifiers that remain only in source text. We present A2RAG, an adaptive-and-agentic GraphRAG framework for cost-aware and reliable reasoning. A2RAG couples an adaptive controller that verifies evidence sufficiency and triggers targeted refinement only when necessary, with an agentic retriever that progressively escalates retrieval effort and maps graph signals back to provenance text to remain robust under extraction loss and incomplete graphs. Experiments on HotpotQA and 2WikiMultiHopQA demonstrate that A2RAG achieves +9.9/+11.8 absolute gains in Recall@2, while cutting token consumption and end-to-end latency by about 50% relative to iterative multihop baselines.

A2RAG: Adaptive Agentic Graph Retrieval for Cost-Aware and Reliable Reasoning

TL;DR

A2RAG tackles the challenge of grounding large language models for knowledge-intensive, multi-hop QA by separating reliability control from progressive evidence acquisition. It combines an Adaptive Control Loop that verifies answers against provenance and triggers bounded rewrites with an Agentic Retriever that starts with local graph evidence and escalates to bridge and global diffusion, finally map-back to source text to recover fine-grained qualifiers. The framework delivers strong evidence-retrieval gains (Recall@2) and substantial efficiency improvements, while remaining robust to extraction loss through provenance-based grounding. This approach offers a practical, cost-aware path to reliable graph-augmented reasoning in production settings, where graph abstractions may be imperfect or incomplete.

Abstract

Graph Retrieval-Augmented Generation (Graph-RAG) enhances multihop question answering by organizing corpora into knowledge graphs and routing evidence through relational structure. However, practical deployments face two persistent bottlenecks: (i) mixed-difficulty workloads where one-size-fits-all retrieval either wastes cost on easy queries or fails on hard multihop cases, and (ii) extraction loss, where graph abstraction omits fine-grained qualifiers that remain only in source text. We present A2RAG, an adaptive-and-agentic GraphRAG framework for cost-aware and reliable reasoning. A2RAG couples an adaptive controller that verifies evidence sufficiency and triggers targeted refinement only when necessary, with an agentic retriever that progressively escalates retrieval effort and maps graph signals back to provenance text to remain robust under extraction loss and incomplete graphs. Experiments on HotpotQA and 2WikiMultiHopQA demonstrate that A2RAG achieves +9.9/+11.8 absolute gains in Recall@2, while cutting token consumption and end-to-end latency by about 50% relative to iterative multihop baselines.
Paper Structure (32 sections, 10 equations, 6 figures, 4 tables, 1 algorithm)

This paper contains 32 sections, 10 equations, 6 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Problem Definition
  4. Related Methods
  5. A2RAG Framework
  6. Overview
  7. Adaptive Control Loop
  8. Entry: Summarized-KB Gating
  9. Exit: Verified Answering via Triple-Check
  10. Iteration: Failure-Aware Rewrite and Bounded Retry
  11. Optional: Human-in-the-Loop Knowledge Base Update
  12. Agentic Retriever
  13. Stage 1: Local Evidence Collection
  14. Stage 2: Bridge Discovery
  15. Stage 3: Global Fallback with Degree-Normalized PPR and Provenance Map-back
  16. ...and 17 more sections

Figures (6)

  • Figure 1: A2RAG Framework Overview
  • Figure 2: Adaptive Control Loop
  • Figure 3: Agentic retriever
  • Figure 4: Stage-wise breakdown of A2RAG's progressive retrieval. Each pie chart reports the fraction of queries that terminate at the local (1-hop), bridge ($K$-hop), or PPR-based global fallback stage, with failed cases shown separately.
  • Figure 5: Robustness to extraction loss on HotpotQA measured by Recall@5 under random KG node/edge deletion.
  • ...and 1 more figures