Neo4j Executive Argues Enterprise AI Hallucinations Need Data-Layer Fixes
Sudhir Hasbe, President and Chief Product Officer at Neo4j, argues that language models produce fluent but wrong answers when they lack the right enterprise facts, so the remedy lies in data architecture rather than prompt tweaking. He notes that plain vector search gave some enterprise RAG teams only about 60 to 65 percent accuracy, while adding a graph filter on metadata and relationships, which he calls GraphRAG, reportedly lifted customers above 90 percent. He also describes context graphs that record the rationale behind decisions, such as why a discount was approved, which systems of record typically omit. The piece presents three ways teams can structure such graphs, with most starting with a semantic map or a unified slice of scattered data. Dan Klein, CTO of Scaled Cognition, is also quoted describing language models as plausibility engines whose outputs can be mistaken for truth.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in