support the new strict mode for openai Structured Outputs

[aviadr1]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

Enhance LangChain by integrating support for this OpenAI's new Structured Outputs mode, this strict Structured Outputs for 100% conformance to developer-supplied JSON Schemas.

support at least one of the following:

Structured Outputs with Function Calling:

Integrate the ability to specify strict: true in function definitions when using OpenAI's function calling feature within LangChain. This ensures that the outputs generated by the model will match the supplied JSON Schema.

Structured Outputs with Response Formats:

Support the new json_schema option in the response_format parameter when using OpenAI's latest models in LangChain. This allows the models to generate responses that adhere to specified JSON Schemas even when not using function calling.

Motivation

generating 100% JSON-schema conforming output is incredibly powerful and will greatly improve the reliability of using LLMs in chains

Proposal (If applicable)

No response

[simonedstrom96]

This would be very useful, awesome if added.

[juan-abia]

Really looking fordward to this!

[Nachoeigu]

I don´t know something regarding this update. Why not to automatically place strict = True, without putting it as a parameter? When is useful to keep strict = False?

[theturtle32]

There's some quite significant overhead when calling OpenAI's endpoints with a new schema for the first time. Their introductory blog post said "Typical schemas take under 10 seconds to process on the first request, but more complex schemas may take up to a minute."

It could also potentially affect backward compatibility, if default behavior changes in people's projects?

[Nachoeigu]

Yes, @theturtle32 . I understood what you mean. But, my question is more focus on: in which cases would I prefer to use strict = False? If I want to use function calling, I always prefer to receive the correct structured output.

[hardchor]

@Nachoeigu I think he kinda already answered your question - you don't just introduce such a "breaking" change (it introduces significant lag on the first request) for everyone by default, even if it improves fidelity. The right way to go here is to expose it as an option.

That being said, I'm looking forward to testing it out soon!

[ShubhamMaddhashiya-bidgely]

I'm looking forward to this feature in langchain.
Is there any workaround until this feature comes into langchain officially?

[ShubhamMaddhashiya-bidgely]

this feature is already released in langchain-openai==0.1.22 🚀

[multib0t]

When I tried switching to langchain-openai==0.1.22 my app was broken. Is there any migration guide?
Is the below still valid and is going to use strict=True in the background once I update.
ChatOpenAI(model='gpt-4o', model_kwargs={"response_format": {'type': 'json_object'}})

[ShubhamMaddhashiya-bidgely]

can you attach the error logs?

[muriloime]

I think this should be closed as per @ShubhamMaddhashiya-bidgely comment

[ShubhamMaddhashiya-bidgely]

yes

[muriloime]

Actually this is implemented only for the function calling, correct?

[arne-rustad]

Can I request that this be implemented for the Azure version of the chat model in this thread, or should I start a new conversation for that?