Natural‑language analytics over enterprise data warehouses is becoming increasingly important, yet production deployments are hampered by hallucinated metrics, invalid joins, wrong grain, unsafe data access, and unsupported explanations. Existing text‑to‑SQL approaches typically ground generation on database schemas or retrieved documentation, but enterprise reporting also demands governed business semantics—approved metrics, dimensions, join paths, filters, and row‑level security.
GROUND (Governed Retrieval Over Unified Normalized Definitions) introduces a framework that constrains LLM‑generated analytics to a governed semantic layer. The system first supplies a unified library of approved definitions, then binds user intent to these governed metrics and dimensions. Before execution, the generated SQL is validated against multiple rules: schema conformity, metric consistency, join legality, grain alignment, filter policies, row‑level security, and cost constraints. If any violation is detected, GROUND retries generation or abstains from execution.
In a synthetic benchmark of 100 enterprise‑reporting questions, GROUND was compared with three baselines—schema‑only text‑to‑SQL, schema‑RAG, and semantic‑only grounding—using the same underlying model. GROUND achieved zero hallucinations across all six evaluated categories, while ungoverned systems breached row‑level security on many queries. Even a semantic‑only condition that provided exact metric definitions but omitted access policies still leaked data, demonstrating that governance cannot be replaced by metric fidelity alone.
The findings were replicated on real U.S. NHTSA vehicle‑safety data with an independently authored gold standard. An adversarial test across four models from three providers showed that GROUND’s enforced guarantees for filters and row‑level security held with zero violations on every model, though judgment‑dependent behaviors such as refusing undefined metrics remained fallible.
Blogger's Review: GROUND proves that embedding a governed semantic layer into LLM‑driven analytics can effectively mitigate hallucinations and protect sensitive enterprise data. Further work on handling ambiguous intents and cost‑aware optimization could broaden its practical impact.