Agent-G: An Agentic Framework for Graph Retrieval Augmented Generation
Keywords: Graph Retrieval-Augmented Generation, Large Language Models, LLM Agents
TL;DR: We propose an agentic framework for use in Graph Retrieval-Augmented Generation.
Abstract: Given two knowledge sources, one containing unstructured documents and the other comprising structured graph knowledge bases, how can we effectively retrieve the relevant information to answer user questions?
While Retrieval-Augmented Generation (RAG) retrieves documents to assist the large language model (LLM) in question answering,
Graph RAG (GRAG) uses graph knowledge bases as an additional knowledge source.
However, there are many questions that require information from both sources, which complicates the scenario and makes hybrid retrieval essential.
The goal is to effectively leverage both sources to provide better answers to the questions.
Therefore, we propose Agent-G, a unified framework for GRAG, composed of an agent, a retriever bank, and a critic module.
Agent-G has the following advantages:
1) Agentic, it automatically improves the agent's action with self-reflection,
2) Adaptive, it solves questions that require hybrid knowledge source with a single unified framework,
3) Interpretable, it justifies decision making and reduces hallucinations, and
4) Effective, it adapts to different GRAG settings and outperforms all baselines.
The experiments are conducted on two real-world GRAG benchmarks, namely STaRK and CRAG.
In STaRK, Agent-G shows relative improvements in Hit@1 of 47% in STaRK-MAG and 55% in TaRK-Prime.
In CRAG, Agent-G increases accuracy by 35% while reducing hallucination by 11%, both relatively.
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3154
Loading