Abstract

Retrieval-augmented generation (RAG) has become the standard approach for grounding generative models in external knowledge. When that knowledge is structured as a graph, GraphRAG methods have emerged to leverage topology to boost retrieval. However, existing approaches predominantly rely on either LLM -based pipelines, which treat graph structure as text to summarise or traverse, or use graph algorithms and neural scoring only as an intermediate step before falling back to document-based retrieval, leaving much of the graph structure unexploited by the generative model. Graph Neural Networks (GNNs) offer a compelling middle ground: natively designed for graph-structured data, learnable end-to-end, and capable of encoding complex relational patterns into compact representations. This talk explores the opportunities and challenges of placing GNNs at the core of GraphRAG pipelines, presenting a GraphNeuralRAG framework across two domains. In multi-hop question answering over knowledge graphs, GNN -based retrieval is shown to address key limitations of document-centric approaches, including recall-precision tradeoffs, token inefficiency, and loss of structural information. In perturbation modelling, PT-RAG (ICLR 2026 Workshop on Generative AI in Genomics) is presented as the first RAG framework for predicting single-cell responses to gene perturbations, followed by a discussion of how gene-gene networks provide a natural graph substrate for extending this into a full GraphNeuralRAG pipeline. The focus throughout is on both what GNNs uniquely enable and what remains hard, offering an honest map of this emerging research direction.