Safety

Text-to-SQL with guardrails

Read-only queries, row limits, schema-grounded columns and an entity map from business terms to columns.

The problem

Letting an LLM write SQL against a production database is risky: it can write unsafe queries, invent columns, or misread business words.

The approach

• SELECT-only queries, and every query gets a row limit.
• The schema (with column descriptions) is the only source of truth for column names.
• An entity-resolution table maps business words (team, leader, service) to the right columns and values.
• Default filters apply business rules, and the agent suggests follow-up questions.

What I learned

• Guardrails belong in both the prompt and the code; never trust only one layer.
• Most wrong answers come from ambiguous business terms, not from SQL syntax.