Just discovered something that feels like a hack? I was working in a spreadsheet with some sales data, exported a bit as CSV. On a whim, I pasted the raw CSV lines into HuggingChat and asked it to write a SQL query to find the top customer by revenue.
It figured out the columns and wrote a proper GROUP BY query! I didn't have to explain the structure at all. 😮
Has anyone else tried using it for this kind of data wrangling? I'm curious how reliable it is for more complex joins or if it makes weird assumptions. This could be a game-changer for my reports.
It's clever for a quick demo, but that's about it. The second you move past a trivial GROUP BY, you'll hit a wall.
Wait until you need to join that CSV snippet against an actual database schema it's never seen. The assumptions it makes about data types, NULL handling, and join keys are often catastrophically wrong, and debugging a subtly incorrect query it wrote is more work than just writing the correct one yourself.
It's a game-changer for generating incorrect reports, maybe.
— skeptical but fair
Your point about assumptions on data types and join keys is critical. I've tested this with multi-table export snippets where timestamp formats differed between CSV extracts, and the LLM defaulted to naive string comparisons that would fail on any actual datetime casting or timezone conversion.
The reliability completely breaks down when dealing with nested JSON structures flattened into CSV columns - it often misinterprets the cardinality, suggesting joins that would produce Cartesian products. For quick exploration it's a useful crutch, but I'd never let it generate production queries without a manual schema validation pass first.