When response_model is optional

[timconnorz]

response_model is currently not an optional parameter, but it would be nice if it was for cases where its unnecessary.
In this case I expect Instructor would skip over the LLMValidator step.

[roodboi]

It should do a straight pass-through if the additional params are not provided - can you share under what circumstances this is not the case?

[timconnorz]

Ah my bad, I was confused by the errors I was getting.

What confused me what the fact that when response_model is specified, you get your response in the exact shape of the model, but when it's omitted, you have to revert back to choices[0].delta and append them to the string yourself.

[roodboi]

Ahh got it - yeah the goal was to ensure full compatibility with the sdk, only adding the enhanced layer when needed.

if all you want is a string response but don't need much else, could you still use the response model with just a simple schema?

eg

Z.object({ content: z.string() })

I almost never use straight sdk at this point, even if just for a single string response - for most cases I usually just want the string parsed out of the response and nothing else - this still gets you that and sets us up to add additional output should you need it.

[jxnl]

I Agree sometimes i use Z.object({ content: z.string() }) since the type of response is often shorter and is less censored.