Are there any plans to replace patching by delegating?

[zby]

For some time we have the openai client object - so instead of patching the library we could just store the openai client object, use it to make calls and then process the response.

The example would become:

instructor = Instructor()
user = instructor.process(
    model="gpt-3.5-turbo",
    response_model=UserDetail,
    messages=[
        {"role": "user", "content": "Extract Jason is 25 years old"},
    ]
)

When creating the Instructor one could also pass a client - this would be useful for tests.

[jxnl]

honestly know, i'm pretty hesitant to introduce a new sdk to wrap an existing sdk. its another layer of abstraction i didn't earn

[zby]

So maybe inheritance? It is like patching but a bit more civilized.
The disadvantage would be losing the "don't mess with me" warning.

But I expect that once you add other LLMs you'll have to make a new api anyway.

[jxnl]

not that interested right now im sure 95% of all llm api calls are either openai or mirror the openai sdk.
if people want to use code the integrates with every single api they can use

https://python.langchain.com/docs/modules/model_io/output_parsers/pydantic
https://www.askmarvin.ai/
https://gpt-index.readthedocs.io/en/latest/examples/output_parsing/openai_pydantic_program.html

[jxnl]

I think inheritance makes sense from the perspective of getting better IDE support for the new word arguments.

but having Instructor()
means i now have to support whatever wide range of key word arguments from any client, rather than letting OpenAI() handle it. azure, openai. mistral. together. etc

you could image class methods

instructor.from_openai
instructor.from_mistral
instructor.from_ollama 

but its a lot of work to support 1% of the user base, esp when each model has different prompts, the library starts being responsible for writing prompts rather than just supporting the function calling and tool calling apis