https://gist.github.com/int19h/4f5b98bcb9fab124d308efc19e530...

Note that in this case it isn't even asked to write specific queries for specific tasks - it's just given one high-level task and the schema to work with.

You're right, though, that the effectiveness of this approach depends very much on schema design and things like descriptive table/column names etc (and even then sometimes you have to make it more explicit than a human would need). You really need to design the schema around the capabilities of the model for best results, which makes it that much harder to integrate with legacy stuff. Similarly, not all representations of data work equivalently well - originally, I gave the model direct access to the typed object graph, and it handles that much worse than SQL. So if your legacy software has a data model that is not easy to map to relational, too bad.

On the other hand, GPT-4 is already vastly better at this kind of task than GPT-3.5, so I think we can't assume that this will remain a limitation with larger models.