Uber's QueryGPT turns plain-English questions into SQL, cutting query time from ~10 to ~3 minutes

What the agent does

An employee asks a question in English. QueryGPT finds relevant tables and example queries with vector search, then writes a SQL query. The user checks and runs it.

How it grew

It began as a proposal at Uber's internal Generative AI Hackdays in May 2023. Later versions split the job into smaller agents: one picks the business area, one picks tables, and one trims columns before writing SQL.

What to copy

• Group your tables by business area. Let the agent pick the area first.
• Store good example queries. They teach the model your naming and join patterns.
• Show the user which tables were chosen, so they can correct it early.

Starter prompt

Written by us, not by Uber.

Business area: {area}
Tables you may use (schemas below): {tables}
Example queries from this area: {examples}

Question: {question}

Write one {dialect} query. Use only the tables above.
Then explain in 3 bullets what it returns and which filters you assumed.