GFM-RAG: Graph Foundation Model for Retrieval Augmented Generation

Retrieval-augmented generation (RAG) has proven effective in integrating knowledge into large language models (LLMs). However, conventional RAGs struggle to capture complex relationships between pieces of knowledge, limiting their performance in intricate reasoning that requires integrating knowledge from multiple sources. Recently, graph-enhanced retrieval augmented generation (GraphRAG) builds graph structure to explicitly model these relationships, enabling more effective and efficient retrievers. Nevertheless, its performance is still hindered by the noise and incompleteness within the graph structure. To address this, we introduce GFM-RAG, a novel graph foundation model (GFM) for retrieval augmented generation. GFM-RAG is powered by an innovative graph neural network that reasons over graph structure to capture complex query-knowledge relationships. The GFM with 8M parameters undergoes a two-stage training process on large-scale datasets, comprising 60 knowledge graphs with over 14M triples and 700k documents. This results in impressive performance and generalizability for GFM-RAG, making it the first graph foundation model applicable to unseen datasets for retrieval without any fine-tuning required. Extensive experiments on three multi-hop QA datasets and seven domain-specific RAG datasets demonstrate that GFM-RAG achieves state-of-the-art performance while maintaining efficiency and alignment with neural scaling laws, highlighting its potential for further improvement.

💡 Research Summary

This paper introduces GFM-RAG, a novel Graph Foundation Model for Retrieval-Augmented Generation, designed to overcome the limitations of conventional RAG systems in capturing complex relationships between pieces of knowledge.

The core problem addressed is that while RAG effectively integrates external knowledge into Large Language Models (LLMs), standard methods retrieve documents independently, struggling with multi-hop reasoning that requires synthesizing information from multiple sources. Although GraphRAG approaches that build explicit graph structures from documents have improved this, their performance is still hampered by noise and incompleteness in the constructed graphs.

GFM-RAG’s framework consists of three main components. First, a KG-index is constructed from the document corpus using Open Information Extraction and Entity Resolution, creating a knowledge graph that serves as a structural index linking entities and their source documents. Second, the central innovation is the Graph Foundation Model Retriever (GFM Retriever). This is powered by a query-dependent Graph Neural Network (GNN). Unlike standard GNNs, it initializes the features of entities mentioned in the user query with the query’s own embedding. It then performs multi-layer, relation-aware message passing across the KG-index to dynamically propagate query-specific information, effectively performing multi-hop reasoning in a single step to identify relevant entities. Third, relevance scores for entities are aggregated via an entity-document inverted index to rank and retrieve the top-K most relevant documents, which are then fed to an LLM for answer generation.

A key achievement is the training of the GFM Retriever as a foundational model. With only 8 million parameters, it undergoes a two-stage training process on a massive dataset comprising 60 knowledge graphs (over 14 million triples) and 700k documents. The first stage is self-supervised pre-training for Knowledge Graph completion. The second stage is supervised fine-tuning for the document retrieval task. This large-scale, multi-task training enables the model to generalize effectively to completely unseen datasets and domains without requiring any further fine-tuning, making it the first graph foundation model applicable out-of-the-box for RAG.

Extensive experiments demonstrate state-of-the-art performance. GFM-RAG outperforms existing baselines (including standard RAG, GraphRAG variants like HippoRAG, and other GNN-based retrievers) on three multi-hop QA datasets (HotpotQA, 2WikiMultihopQA, MuSiQue). Crucially, it also achieves superior results on seven diverse domain-specific RAG datasets (e.g., biomedical, customer service), showcasing its remarkable generalizability. Furthermore, the model is shown to be efficient, replacing costly iterative retrieval-reasoning loops with a single-pass operation, and its performance aligns with neural scaling laws, indicating potential for further improvement with increased scale.

In summary, GFM-RAG successfully bridges the gap between graph-based reasoning and foundation models for RAG. By leveraging a pre-trained, query-dependent GNN on a structurally indexed knowledge graph, it achieves superior, efficient, and highly generalizable retrieval for complex knowledge-intensive tasks, paving the way for more robust and intelligent knowledge-enhanced language models.