GROUND Framework Reduces LLM Hallucinations in Enterprise Analytics

Large Language Models (LLMs) are increasingly being used for natural-language analytics over enterprise data, but their deployment in production is hindered by issues like hallucinated metrics, incorrect data joins, and security vulnerabilities. Current text-to-SQL systems often rely on database schemas or documentation, which lack the crucial governed business semantics—approved metrics, dimensions, and security policies—that enterprise reporting demands. This paper introduces GROUND, a novel framework designed to address these challenges. GROUND integrates a governed semantic layer, providing LLMs with approved definitions and validating generated SQL against a comprehensive set of rules including schema, metric, join, grain, filter, security, and cost. If violations occur, GROUND either retries the generation or abstains. In a benchmark using synthetic enterprise reporting questions and real-world data, GROUND demonstrated zero hallucinations across all evaluated categories and successfully enforced row-level security, unlike ungoverned systems. This highlights that robust governance, beyond just metric fidelity, is essential for secure and reliable LLM-based enterprise analytics.