ISTI-TALK – Graph-NeuralRAG: On the Opportunities and Challenges of GNNs for GraphRAG

Day - Time: 03 June 2026, h.12:00

Place: Area della Ricerca CNR di Pisa - Room: C-29

Speakers

• Andrea Giuseppe Di Francesco

Referent

Davide Rucci

Abstract

Retrieval-augmented generation (RAG) is the standard approach for grounding generative models in external knowledge. When that knowledge is structured as a graph, existing GraphRAG methods either reduce structure to text or use it only for intermediate retrieval, leaving it unexploited at generation time. We propose Graph-NeuralRAG, a framework placing GNNs at the core of GraphRAG. We show that injecting GNN embeddings as soft tokens into a frozen LLM causes representations to converge to a near-constant vector regardless of the query — a phenomenon we term token collapse. We trace this to an embedding norm mismatch with the LLM's token space and connect it to the low intrinsic rank of soft prompts observed in prior work. To address this, we propose a two-stage training strategy combining contrastive alignment over gold entities with projection fine-tuning. We conclude by discussing open challenges for integrating GNNs into RAG pipelines.