The Stack Overflow Podcast Connecting the dots for accurate AI
13 snips
May 12, 2026 
Philip Rathle, CTO at Neo4j, a leader in graph databases and knowledge layers for AI. He explains why model-only approaches struggle in enterprises. He describes Graph RAG: combining knowledge graphs with vectors for targeted, connected context. He discusses avoiding context rot, deterministic multi-hop reasoning, graph embeddings, and real-world production uses like Uber and Walmart.
AI Snips
Chapters
Transcript
Episode notes
Models Alone Are Incomplete For Enterprise Agents
- Relying only on an LLM for agent decisions is limited because models are stale, stochastic, and lack judgment.
- Philip Rathle argues agents need a live context layer with latest data and deterministic reasoning to be safe for enterprise use.
Add A Knowledge Graph To RAG
- Do not stop at vector RAG; add a knowledge graph to your retrieval pipeline to improve accuracy and explainability.
- Philip recommends calling out to a knowledge graph (graph RAG) so context is targeted and connected across silos.
More Context Can Make LLMs Worse
- Larger prompt context can worsen LLM answers due to context rot; being surgical with context improves outcomes.
- Philip cites a meta-study on 1,500 RAG papers and recommends zeroing in on relevant entities a level or two out in the graph.