[kiitos]

Huh, weird stuff!

Your linked image shows what appear to be tables, and the little arrows appear to represent entity-relationships between them. But I'm not sure how you'd get useful DDL out of it -- none of the columns have types, no indices, etc.!

Maybe an LLM could sketch out a DDL skeleton from a picture, which someone could use as a starting point?

[Philpax]

LLMs can infer a lot of details from "common sense" (i.e. statistical association). As a human, I can figure out what each field's types should be, so your frontier LLM could, too: https://chatgpt.com/share/680e3a90-8660-8003-998f-91ad98e32f...

[kiitos]

You and a LLM can guess at what the types might be, but those guesses are a suggestion that a human needs to evaluate, they're not something you can just pass thru as assumptive defaults. In your link, for example, I would definitely not want CategoryID to be an INT, or UnitPrice to be a DECIMAL, or etc.

[Philpax]

Sure, yeah - but just like with a human, I can provide additional domain context that can clarify its answer. I see your point - you need to know what to provide in order to get the result you want - but I think that today, that makes it a very useful tool, and tomorrow, it'll be able to make those clarifications itself.

(Out of curiosity, what would you use instead? I'd default to INT/DECIMAL respectively myself - would love to know what your thinking is here!)

[kiitos]

I'd default to CategoryID being a string, with uniqueness constraints defined on the categories table. And UnitPrice to be a composite type, containing a [u]int amount and a string denomination, e.g. `10000 USD_CENT` or something like that.

But what you or I or anyone thinks is the right type for these or any columns, is totally beside the point. The point is that the type for a column isn't really assume-able by an LLM, at least not automatically.