Graphs move from niche database to enterprise knowledge layer for AI systems
SiliconANGLE Kelly Knight ● Covered by 3 sources
Neo4j's CTO says AI is settling on a standard fix for LLM hallucinations: knowledge graphs instead of just vector search. A UK study found the graph approach makes AI agents way more accurate and efficient.
Based on reporting by SiliconANGLE, Kelly Knight — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Four years on from ChatGPT's debut, the enterprise AI conversation has quietly shifted. According to Philip Rathle, chief technology officer at Neo4j, companies have stopped tinkering and started converging on a common blueprint for keeping large language models honest: something people are now calling the enterprise knowledge layer. Speaking with theCUBE's John Furrier at Neo4j's GraphTalk event, Rathle framed this layer as the place where an organization's data, structure and agent memory actually live — outside the model, not baked into it.
The pattern underneath this is GraphRAG, where an LLM reaches out to a knowledge graph rather than relying purely on whatever it has memorized or pulled from a vector database. Rathle's pitch is simple: externalize the knowledge, and you get accuracy, explainability and governance that a model alone can't offer. Neo4j has been saying this for a while based on customer experience. What's new is that outside researchers are now backing it up.
The UK's National Innovation Centre for Data ran a comparison between GraphRAG and vector-only retrieval, the two dominant approaches to making AI agents more reliable. GraphRAG came out well ahead — agents were 80% more truthful, answered more than twice as many questions correctly, and did it while burning fewer tokens. That's not a marginal edge. It's the kind of gap that turns a skeptical CIO into a believer, especially once token costs and error rates start showing up on a budget line.
Rathle's favorite proof point is a tax agency engagement. Once the agency's data got modeled as a graph instead of sitting in flat tables, the team spotted more than $100 million in tax fraud within 48 hours of starting a proof of concept. His explanation is almost blunt: teams staring at relational data are working with blinders on, and the graph view suddenly makes patterns that were invisible in two dimensions blindingly obvious.
What ties all this together is a maturing industry finally agreeing on where the line between model and knowledge should sit. The knowledge doesn't belong inside the LLM's weights, where it's opaque and unaccountable. It belongs in a system enterprises can inspect, govern and update independently — and increasingly, that system looks like a graph.
My take — AI-written commentary, not fact-checked reporting
The tax fraud example is the kind of number that should make procurement teams sit up, but it's worth remembering this came from a paid Neo4j event, with a Neo4j executive citing a Neo4j-friendly outcome — context that doesn't make the underlying academic comparison less real, just worth separating from the marketing. Still, the 80% truthfulness gap from an independent UK research body is a genuinely useful data point in a field drowning in vague accuracy claims. Enterprises chasing AI reliability would do well to treat graphs as infrastructure rather than a niche database choice, because flat tables clearly aren't cutting it once agents start making decisions instead of just answering trivia.
Read more about this at: SiliconANGLE