Unlocking Multi-Hop Reasoning with GraphRAG

Session track: AI Engineering

Session time: 05:00 am - 05:30 am UTC November 12

Session description:

Retrieval-Augmented Generation (RAG) has become a standard pattern for grounding large language models—but most implementations rely solely on vector similarity search. While useful, this approach often misses the most important aspect of knowledge: relationships. In this session, we explore how to build a GraphRAG system using Neo4j, where context is not just retrieved—but connected. We’ll design a graph-native architecture that combines a domain graph (events, speakers, sessions) with a lexical graph (documents, chunks, embeddings), and show how linking them unlocks richer, explainable AI responses. Using real-world scenarios, you’ll see how graph traversal strategies—such as breadth-first and depth-first search—enable multi-hop reasoning that traditional vector RAG cannot achieve. Instead of returning isolated chunks, GraphRAG traverses relationships to surface deeper insights, including attribution, context, and cross-document connections. We’ll also dive into how Cypher-powered retrieval complements vector search, allowing you to move from “similar text” to precise graph queries. You’ll learn when to rely on embeddings, when to traverse the graph, and how to combine both effectively within Neo4j. Finally, we extend the architecture using an agentic AI layer, where intelligent agents dynamically decide: When to use vector similarity vs graph traversal How to decompose complex, multi-part questions How to generate and execute Cypher queries for structured retrieval This results in a flexible, production-ready system capable of handling both exploratory and deterministic queries. By the end of this session, you will learn: How to model and connect lexical + domain graphs in Neo4j How to implement GraphRAG with traversal strategies How to leverage Cypher for structured, explainable retrieval How to orchestrate retrieval using agent-based patterns