Beyond RAG and Fine-Tuning
Schema-Grounded SQL Generation for Verifiable Enterprise Analytics
Smart Data staff
Enterprises want large language models working across their structured data, answering real analytics questions, without hallucinated numbers or compute that grows on its own. The standard approaches, retrieval-augmented generation and fine-tuning, were built for documents, not schemas. This talk shows how grounding generation in the schema makes every answer checkable and keeps the cost of asking visible.
- 1
Where RAG and fine-tuning break down
Similarity search and trained weights both miss what gives a table its meaning: keys, grain and join paths.
- 2
Schema grounding makes output verifiable
Constrain generation to the catalog, then check every table, column and join before the query runs.
- 3
Compute costs stay in view
Estimate, cap and cache at the same checkpoint, so usage cannot grow the warehouse bill quietly.
Given at the Ohio AI Conference in Dayton on Aug 26, 2026, this talk looks at getting analytics out of structured enterprise data with an LLM. Instead of answering from retrieved text or a fine-tuned model, the model writes SQL against the enterprise schema, and the database produces the answer.
Because every answer comes from a query, anyone can check it by reading or rerunning that query. The talk compares the approach with retrieval-augmented generation and fine-tuning for structured data, and covers keeping hallucination and compute costs in check.