Knowledge graph architecture gives enterprises ownership of the AI intelligence they create

Shan Rizvi has a blunt message for anyone building enterprise AI around retrieval-augmented generation: you're solving the wrong problem. RAG is great at fetching relevant chunks of text, but fetching isn't reasoning. It can't trace how it got to an answer, and it doesn't get smarter the more you use it. Rizvi, founder of Thumos Care and a self-described context architect, told theCUBE's John Furrier that the fix is structural, not algorithmic — enterprises need knowledge graphs, not smarter search.

His go-to example is oddly mundane and all the more convincing for it: a startup building AI to find substitute truck parts. Picture needing every replacement part from manufacturers that satisfy several conditions at once, buried somewhere inside millions of SKUs. Search chokes on that kind of multi-hop, relational puzzle. A graph doesn't. It walks the connections, backs up, tries another path, and keeps recursing until it lands on something solid — which, Rizvi points out, is basically how people reason through an argument: premise, logic, conclusion.

None of this works if latency blows up, though, and Rizvi learned that the hard way building a medical knowledge graph for Thumos Care on Neo4j, linking diseases, symptoms, drugs and clinical guidance. Not every graph database handles tens of millions of nodes without choking, and not every query needs the full recursive treatment. Simple questions should get fast, single-hop answers; the deep traversal only kicks in when the problem actually demands it. Skip that tiering and you end up with a system too slow to ship.

Rizvi also draws a sharper line than most between what belongs in the graph and what doesn't. Fast-changing structured data — inventory numbers, transaction records — should stay put in its native database, with only its shape mirrored in the graph so an AI agent knows where to look. Unstructured stuff like Slack threads, meeting transcripts and support logs is the real payoff: pull it directly into the graph, tie it to the right entities, and it becomes durable context instead of a pile of forgotten messages.

The bigger claim underneath all this is about ownership. Rizvi argues that data sitting in a database is just data — replaceable, commoditized. Structure it into an ontology that captures how your domain actually thinks and decides, and it becomes intelligence that compounds every time someone uses it. That, he says, is the real divide opening up in the agentic AI era: companies building disposable search tools versus companies quietly building something they'll actually own.