What is text2sql?

Text2SQL lets you ask a question about a PostgreSQL database in plain English and get back an answer, while trying hard to make sure the automatically generated SQL cannot damage the database or return a misleading result. Turning a natural language question into a database query with an AI model is easy to demonstrate but risky in practice, because even a model that works most of the time can still occasionally produce a destructive command or a confidently wrong answer. This project's approach is to add several independent safety checks rather than simply trusting the model. A question passes through several stages before it becomes a final response. First, a simple keyword check looks for destructive intent in the plain English question itself, before any AI call happens. Next, an AI model, primarily one from Groq with a Google Gemini model as a backup, generates the SQL along with an explanation and a confidence score, and this generated SQL is checked for correct syntax with one retry allowed. The SQL is then analyzed structurally to confirm it is a single read-only SELECT statement, is not accessing tables or columns outside an approved list, does not call known dangerous database functions, and has a limit on how many rows it can return. Even after all of that, the query only ever runs against a database account that has read-only permission and a time limit, so a write operation would be blocked at the database level even if every earlier check somehow failed. Finally, a second, independent AI call reviews the question, the SQL, and a sample of the results to judge whether the answer actually addresses what was asked. The project is built as a monorepo with a Cloudflare Workers backend using the Hono framework, a Next.js frontend, and a serverless Postgres database accessed through the Drizzle database toolkit. Running it locally requires installing dependencies with pnpm, setting up a database, and providing API keys for the Groq and Gemini AI services. The README notes some current limitations: it only works with one fixed demo database schema of customers, products, and orders, it is built for a single set of database credentials rather than multiple users, and the final AI-based result check is not a mathematical proof that the SQL is correct, just an added layer of scrutiny. The README does not state a license.