# Don't Forget to Connect! Improving RAG with Graph-based Reranking

**Authors:** Jialin Dong, Bahare Fatemi, Bryan Perozzi, Lin F. Yang, Anton Tsitsulin  
**arXiv:** [2405.18414](https://arxiv.org/abs/2405.18414)  
**Date:** 2024-05-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2405.18414v1/

## Key points

- Authors: Jialin Dong, Bahare Fatemi, Bryan Perozzi, Lin F. Yang, Anton Tsitsulin
- Focus: Enhancing performance of RAG responses by grounding generation with context from existing documents
- Handling documents with incomplete information
- Reasoning about connections between documents effectively
- Solution: G-RAG reranker leveraging graph neural networks (GNNs) for establishing connections within the RAG framework
- Incorporates document-to-document connections and semantic information through Abstract Meaning Representation graphs
- Outperforms existing state-of-the-art approaches while maintaining a smaller computational footprint
- PaLM 2 significantly underperforms compared to G-RAG as a reranker
- Highlights the critical role of reranking in optimizing RAG outcomes, even with Large Language Models.

## Summary

In their paper titled "Don't Forget to Connect! Improving RAG with Graph-based Reranking," authors Jialin Dong, Bahare Fatemi, Bryan Perozzi, Lin F. Yang, and Anton Tsitsulin delve into the realm of  and its advancements in enhancing the performance of  responses. The core focus lies in grounding generation with context from existing documents, particularly addressing scenarios where documents contain partial information or have less obvious connections to the context at hand. The researchers aim to tackle two fundamental questions surrounding RAG generation: how to handle documents with incomplete information and how to reason about connections between documents effectively. To address these challenges, the authors introduce , a reranker that leverages graph neural networks (GNNs) to establish connections between the retriever and reader within the RAG framework. By incorporating both document-to-document connections and semantic information through Abstract Meaning Representation graphs, G-RAG serves as a context-informed ranker for RAG systems. Notably, their method outperforms existing state-of-the-art approaches while maintaining a smaller computational footprint. Furthermore, the study evaluates the performance of PaLM 2 as a reranker and finds it significantly underperforming compared to G-RAG. This outcome underscores the critical role of reranking in optimizing RAG outcomes even when utilizing Large Language Models. Overall, this research sheds light on the importance of refining retrieval augmented generation processes through innovative techniques like graph-based reranking for more effective contextual understanding and response generation in natural language processing tasks.