GraphRAG: Using LLMs to Build and Interrogate a Regulatory Knowledge Graph
summary
GraphRAG combines retrieval-augmented generation with a Neo4j-backed regulatory knowledge graph to return not just passages but the metadata and links between them, reducing hallucinations and adding nuance. It powers a logic graph covering 19 source acts, 138 sections and 987 modeled questions, uses embeddings and a vector index in Neo4j AuraDB, extracts cross-references by generating Cypher via few-shot and fuse-shot learning, and employs a “target similarity generation” loop that compiles and iterates generated logic code until it runs — a workflow reported as effective at expanding and interrogating policy datasets.
AI-written summary · about
chapters (8)
- 0:00Introduction to GraphRAG
- 3:00LLMs and hallucinations
- 5:50What is GraphRAG
- 10:04Building a knowledge graph
- 11:09Extracting references
- 14:22Target similarity generation
- 19:48Interrogating the knowledge graph
- 24:45Q&A
related talks
39mEnhancing RAG with Neo4j Knowledge Graph
25mAdaptive GraphRAG: A Framework for Knowledge Graph Quality, Consistency, Evolution
16mEnhancing Retrieval-Augmented Generation with GraphRAG Patterns in Neo4j
21mAgentic GraphRAG: Autonomous Knowledge Graph Construction and Adaptive Retrieval
27mGraphReader Agentic RAG: Rethinking Long-Context Retrieval Systems
12mAgentic GraphRAG: Autonomous Knowledge Graph Construction and Adaptive Retrieval
19mBeyond Retrieval: Adding a Reasoning Layer to RAG
30mAgentic RAG with Knowledge Graphs Explained