Are Your Semantic Models AI-Ready? What Knowledge Graphs Teach Us About Building Context for LLMs

Session track: Data Intelligence

Session time: 08:00 pm - 08:30 pm UTC November 12

Session description:

Every enterprise data team racing to deploy LLM-powered analyst tools, Snowflake Cortex Analyst, Databricks Genie, and others, hits the same wall: the SQL the model generates is wrong more often than anyone expected. The data is there but the context isn't. The knowledge graph community has been solving this problem for decades. Typed entities, explicit relationships, and disambiguated semantics are exactly what LLMs need to reason reliably and exactly what most enterprise semantic models lack. In this session, Nivedita will share findings from Semantic Model Inspector, an open-source tool she built to evaluate the AI-readiness of enterprise semantic models. She will walk through the specific failure modes that break LLM reasoning like ambiguous entity definitions, missing join semantics, underspecified metrics and show how knowledge graphs natively solve each one. She will then walk you through a working pattern: taking a public semantic model, encoding the same domain in Neo4j as a typed knowledge graph, and exposing that graph context to an LLM before SQL generation. You will see the graph schema, the Cypher patterns that capture what warehouse-native semantic layers miss, and side-by-side code examples of LLM query generation with and without graph-backed context. You will learn a taxonomy of semantic gaps that break LLM reasoning, a graph-native pattern for closing them, and how to decide when a semantic layer is enough and when a knowledge graph earns its place in your stack.