From Vector RAG to GraphRAG: Building Context Graphs as Durable Memory for Production AI Agents

Session track: AI Engineering

Session time: 07:30 pm - 08:00 pm UTC November 12

Session description:

Most conversational agents begin with vector RAG and quickly hit a wall: semantic search retrieves similar text chunks but misses the connected facts, relationships, and provenance that agents need to reason reliably. In this session, Maulik Bhatt will show how modeling enterprise knowledge as a Neo4j graph transforms brittle retrieval into durable, explainable agent memory. Drawing on his experience architecting AWS Quick Suite's Conversation Service (310,000+ accounts, 8M daily messages), he will walk through evolving a production RAG pipeline into GraphRAG with live Cypher and graph data models. You will learn how to construct a knowledge graph from unstructured and structured sources using LLM-assisted entity and relationship extraction; how to combine Neo4j vector indexes with multi-hop graph traversal for grounded, complete retrieval; and how to persist conversation state and provenance as a context graph that gives agents continuous, auditable memory instead of stateless prompts. He will also share production lessons on schema design, keeping the graph fresh from streaming updates, and latency tuning along with the retrieval-quality wins and pitfalls measured when moving from vector-only to GraphRAG. You will leave with concrete graph data models, reusable Cypher retrieval patterns, and a clear blueprint for using Neo4j as the memory and context layer for reliable, explainable production AI agents.