Oracle Deep NL2SQL treats text-to-SQL like code, with gold-query regression tests

Text-to-SQL, where you ask a question in plain language and a model writes the query, usually fails in a quiet way: the SQL runs, but it answers the wrong question. Oracle’s engineering post on Deep NL2SQL, published during its AI World event, describes an agent built around that problem.

How it works

• A business knowledge graph. Schemas, column notes, sample questions, documents and business definitions are organized into a hierarchy of topics, so a definition set at a parent topic applies to everything beneath it. For each question, vector, text and graph search pull only the relevant topics into the model’s context.
• It asks instead of guessing. If the knowledge does not settle what “active customer” or “last quarter” means, the agent asks the user a clarifying question.
• Plan, run, review. The agent breaks the question down, explores the data, writes and runs SQL, checks its own result, and returns the SQL with its reasoning. Users can vote answers up or down, and confirmed answers go back into the graph.

The part worth copying

Changes to the knowledge graph are treated like code changes. Teams register a “gold” set of natural-language questions with vetted SQL, and the gold set is rerun whenever the graph changes. Edits stay unpublished until those answers still match. It is a regression test suite for a semantic layer, and the idea works with any text-to-SQL stack.

Oracle says users can get “close to 100% accuracy”, but only after the upfront work of building the knowledge base and gold queries. It publishes no benchmark for that, and the post gives no availability date or pricing.

The wider Oracle release

The same day, Oracle announced a Business Semantic Layer and an agent-driven AI Data Catalog for its Autonomous AI Lakehouse, both due “within the next 12 months”. The semantic layer defines dimensions, measures and KPIs once and feeds them to Select AI, NL2SQL and retrieval. Oracle also listed SQL create, update and delete on Apache Iceberg tables, shared access control across Oracle and other data, and a Data Lake Accelerator it says runs up to 3x faster on a 100 TB TPC-DS test of its own. InfoWorld notes Oracle gave no dates for several of these.