Paper Reading Night by HydraDB: Graph Neural Retrieval for LLM Reasoning

​This Week’s Paper

​GNN-RAG: Graph Neural Retrieval for Large Language Model Reasoning

​Most RAG systems retrieve text passages. But what happens when answering a question requires reasoning across multiple entities, relationships, and hops in a knowledge graph?

​This week, The Graph Club will explore GNN-RAG, a retrieval architecture that combines graph neural networks with large language models. The GNN identifies relevant entities and multi-hop reasoning paths within a knowledge graph, then converts those paths into context that an LLM can use to generate an answer.

​The paper reports SOTA performance on WebQSP and Complex WebQuestions, with particularly strong results on multi-hop and multi-entity queries. A key takeaway: a tuned 7B model paired with better retrieval and structured reasoning can match or outperform GPT-4 on these benchmarks.

​What We’ll Discuss:

• ​Why conventional vector retrieval struggles with relational and multi-hop questions

• ​How GNN-based retrieval differs from embedding-based and graph traversal approaches

• ​How reasoning paths are selected and converted into LLM context

• ​Where GNN-RAG performs well, where it breaks down, and what it means for production AI systems

• ​Whether better retrieval architecture can reduce dependence on increasingly large models

​The discussion will be led by Aditya Srivastava, AI Researcher & Engineer at HydraDB and an alumnus of Carnegie Mellon University and IIT Bombay.

​Who Should Attend?

​AI researchers, ML engineers, data and infrastructure engineers, graduate students, and builders working on RAG, knowledge graphs, agent memory, graph machine learning, or LLM reasoning.

​You do not need to be an expert on the paper. Come prepared to question the methodology, challenge the conclusions, and discuss how the ideas might translate into real systems.

​About The Graph Club

​The Graph Club is a paper-reading community for researchers and engineers who want to stay close to emerging AI research, strengthen their understanding of foundational ideas, and connect academic work to the systems being built today.

​About HydraDB

​HydraDB is a fast graph database built on object storage, purpose-built for modern AI workloads (think agent memory, company brain, ontologies etc.). We make AI stateful.

​Read the paper before the session:
https://arxiv.org/html/2405.20139