PyData London

Register: [https://www.meetup.com/pydata-london-meetup/events/315916503/ ](https://www.meetup.com/pydata-london-meetup/events/315916503/ )

Sefik Serengil - RAG Is Dead, Long Live GraphRAG!

Traditional RAG pipelines typically break documents into independent chunks—such as paragraphs or pages—and convert each chunk into an embedding. While effective for basic retrieval, this approach loses the semantic continuity that naturally exists across chunk boundaries. A sentence at the end of one page may be tightly connected to the first sentence of the next, yet classic RAG treats them as unrelated vector islands. Neo4j’s open-source GraphRAG project addresses this fundamental limitation by converting chunks into a graph representation where entities, relationships, and contextual links are explicitly modeled. Instead of isolated embeddings, GraphRAG creates a connected structure in which both the chunks themselves and the entities inside them can be embedded using graph algorithms and LLM-based backends. This allows semantic meaning to flow across the graph, preserving context at multiple levels while enabling richer retrieval and more accurate reasoning.
In this talk, we will explore how GraphRAG builds these graph-native representations, how it leverages Neo4j to store and traverse them, and how its embedding strategies differ from conventional RAG systems. We will also discuss practical workflows, example architectures, and the benefits of using a graph-oriented approach for tasks that require higher-fidelity context understanding. Attendees will gain a clear picture of why GraphRAG matters, what problems it solves, and how to start using the open-source package in real-world applications.