RAG at Scale: Logging, Traceability, and the Architecture for Control
summary
Live production RAG uses Neo4j to store 15,000 chunks (512-byte chunks with ~20–30 overlap), builds kNN edges, runs kNN + Louvain community detection, and measures pairwise similarities (one cluster of 17 chunks has avg similarity ~0.99 and 272 edges). The system collapses near-duplicates into a single node while preserving source URLs, logs prompt→response→chunk lineage, surfaces anomalous long texts and outdated docs, maps conversation paths across clusters, and gives actionable curation signals that demonstrably reduce noise and expose blind spots.
AI-written summary · about
chapters (7)
- 0:00RAG and graph databases
- 3:31Knowledge graph for RAG
- 8:54Cluster curate philosophy
- 15:20RAG datasets at scale
- 21:44Tracking LLM interactions
- 26:42Conversation patterns feedback
- 30:11RAG improvement summary
related talks
39mEnhancing RAG with Neo4j Knowledge Graph
27mOptimizing RAG Systems in Practice: Challenges and Proven Solutions
56mRAG: from dumb implementation to serious results
25mAdaptive GraphRAG: A Framework for Knowledge Graph Quality, Consistency, Evolution
13mOptimizing Real-Time Job Matching: Integrating Neo4j With RAG
30mAgentic RAG with Knowledge Graphs Explained
22mThe Power of RAG at Scale: A Year of Explosive Growth and Observab...
42mTo RAG or not to RAG