Fix model construct

[guillaq]

Found an issue with autopilot where we were streaming dictionaries instead of pydantic objects when the output had nested objects.

The issue was undetected until now because:

• we don't have nested objects when streaming in our internal use
• autopilot did test with nested objects but because we did not have type annotations until https://github.com/WorkflowAI/autopilot/pull/14, the code was using a dictionary directly

[yannbu]

do you want to check ingredient.name == "salt" ?

[yannbu]

So models[1] is a NestedModel without the "field2" attribute?

That would mean that accessing NestedModel.field2 on the chunk processing will raise an AttributeError. And this error can be predicted by the type check... So that will crash at runtime.

I don't have the perfect solution for that. Maybe we should enforce all agent output classes to only have optionals when streaming in the newest SDK versions ? If not raise an error even before starting the stream ?

This way field2 will be None instead of absent; and you are forced by the type checker to handle the case where "field2" is None in the stream processing...

[guillaq]

We mention the caveat of non optional fields in the Readme.
There is no way to do "partial validation" in Pydantic unfortunately.

Another way to handle the issue would be to compute "healthy defaults" for each field but it might get quite complex.. I'll make a quick attempt.

[yannbu]

Ok 👌, thank you!
Having warnings when streaming an output with non-optional would be useful too.

[yannbu]

thanks @guillaq, one important comment, I think model_construct is a dangerous function because it makes type checking completely useless :/

Also, do we have an issue to update the SDK version for our API?

[guillaq]

@yannbu I managed to create an optional version of pydantic models to allow for partial validation of payloads.
Basically like a "simple" version of https://github.com/team23/pydantic-partial/tree/main

I'll check autopilot with it.

[guillaq]

Also, do we have an issue to update the SDK version for our API?

We do not