What's the best way to check if an openai.Client() was patched by instructor?

[gustavocidornelas]

I want to know if there's a good way to programmatically check if an openai.Client() has been patched by instructor or if it is the "original" OpenAI client.

For example:

import instructor
import openai

openai_client = openai.Client()
# openai_client is the original OpenAI client -> so its chat.completions.create method will return a `ChatCompletion` object

patched_client = instructor.patch(openai_client)
# patched_client was patched by instructor -> so its chat.completions.create method might return an object defined by `response_model`

I'd like to know what's the best way to write a function like is_patched_client(client: openai.Client) -> bool, such that:

is_patched_client(openai_client)
# returns False

is_patched_client(patched_client)
# returns True

I can obviously check the type of the object I get from chat.completions.create, but ideally, I'd like to know if the client was patched or not just based on the client itself.

Any help is appreciated!

[jxnl]

this will be possible in 1.0 if you can wait a week or so! It'll be

type(client) == InstructorOpenAI

[gustavocidornelas]

kk, thanks! In the meantime, I'll keep checking the output type.

Also, I'd be great to see some instructor-related metadata as part of the response model, such as the number of retries that actually happened, the errors (stack traces), etc. I'm just thinking out loud now, I haven't checked if that's already possible or not :)

[jxnl]

we do that as part of logging

where would you want to see it all? like

resp = create()
resp._metadata? 

[gustavocidornelas]

Yes, something like that. As part of resp, continue to be able to access resp._raw_response and additionally have resp._metadata where I could see some instructor-specific metadata.

[gustavocidornelas]

The way I'm thinking about it is:

• If I have a lot of retries, it means that the model is having a hard time conforming to the structure I expect.
• The stack trace tells me specifically which fields from the response model were particularly challenging.
• Given my interest in minimizing the number of retries to keep the system's response time low, the stack traces tell me exactly where my prompt needs polishing (either via Pydantic or via the regular messages prompt).
• Looking at the logs is not ideal if I want to do this at scale. If I can access this metadata as part of the response, I can monitor and test it.

[jxnl]

sounds good, we'd have to keep it as BaseModel

since

response_model=bool will be impossible to attach data too without giving you some wrapper hell.