GraphRAG with Neo4j+GNN+LLM

LLMs are powerful, but they struggle with structured reasoning and private knowledge. GraphRAG bridges the gap by combining graph databases, graph algorithms and machine learning to enhance retrieval and reasoning.
Join us to explore Neo4j, GNNs and LLMs in action, see cutting-edge techniques, and connect with fellow developers & researchers. No prior GraphRAG experience is needed - just curiosity!

GraphRAG with Neo4j+GNN+LLM.
Modern LLMs excel in many tasks but struggle to reason over large-scale graph-structured data and cannot connect with private knowledge. Given any question towards a closed dataset, GraphRAG is a powerful approach that retrieves relevant subgraphs and prompts them as context to an LLM. This approach combines many techniques from retrieval-augment generation (RAG) and knowledge-base question-answering (KBQA).

In this meetup, we’ll explore how the use of graph databases, graph algorithms and graph machine learning can together enhance GraphRAG performance beyond the state-of-the-art.
We will present two methods:

• A modular framework that uses Cypher and Graph Data Science algorithm to retrieve from Neo4j and a finetuned GNN+LLM for reasoning.
• A new approach that finetunes a Text2Cypher model with constrained decoding that generates provably correct and optimal Cypher for retrieval.

We will also show the latest and upcoming features in Pytorch Geometric (PyG) that relate to GraphRAG using GNNs and Graph Transformers on NVIDIA CUDA.

Basic knowledge of Neo4j, Cypher and working with open-source LLMs is assumed but no expertise of GraphRAG is required. The presentation will be suitable for both software engineers and researchers.