Text-to-SQL failures stem not from poor SQL generation but from LLMs lacking knowledge of what your data means. The solution is to build a canonical knowledge layer first — a spec combining a data model, taxonomy (what metrics exist and what's out of scope), and ontology (what relationships and joins actually compute). Evidence from Anthropic's internal analytics stack and controlled studies shows that providing this meaning layer raises accuracy from ~21% to 95%+ and adds 17–23 points across all tested models. The key insight is to write definitions before building the model, then use the same spec to both generate the data model and answer queries — keeping the two in sync automatically. This approach mirrors how GraphRAG works over unstructured text, but applied to structured data. dltHub is building tooling (AI Workbench) that implements this workflow, reportedly running 20–50x faster than manual modeling.
