Z.ai models

[alexeygrigorev]

Description

Z.ai provides OpenAI-compatible chat completions interface.

This is how I currently use it:

import os

from pydantic_ai.models.openai import OpenAIChatModel
from pydantic_ai.providers.openai import OpenAIProvider

class ZaiChatModel(OpenAIChatModel):
    def _process_response(self, response):
        response.object = 'chat.completion'
        return super()._process_response(response)

zai_model = ZaiChatModel(
    model_name='glm-4.5',
    provider=OpenAIProvider(
        api_key=os.getenv('ZAI_API_KEY'),
        base_url='https://api.z.ai/api/paas/v4/'
    )
)

The reason I need to wrap it into ZaiChatModel is that they don't set "object" to "chat.completion" and the validation fails because of that

So I was wondering what's the best way to use it with PydanticAI? Is there a better way of dealing with this problem? Also, would it be something worth including in the framework?

References

No response

[DouweM]

@alexeygrigorev Ideally they'd add the field to their API responses if they claim to be Chat Completions API compatible :) Do you have a way of getting that feedback to them?

If that's not an option, we could set the missing field explicitly before we do validation, like we already do if the created timestamp is missing:

pydantic-ai/pydantic_ai_slim/pydantic_ai/models/openai.py

Lines 464 to 468 in fa68251

	if response.created:
	timestamp = number_to_datetime(response.created)
	else:
	timestamp = _now_utc()
	response.created = int(timestamp.timestamp())

[TerenceLiu98]

I also tried the GLM with OpenAIChatModel

when, running request it shows:

  File "/home/terenceliu/srv/XXX/.pixi/envs/default/lib/python3.13/site-packages/pydantic_ai/models/openai.py", line 522, in _process_response
    raise UnexpectedModelBehavior('Invalid response from OpenAI chat completions endpoint, expected JSON data')
pydantic_ai.exceptions.UnexpectedModelBehavior: Invalid response from OpenAI chat completions endpoint, expected JSON data

p.s. Z.ai also provides Anthropic style API, which is fine with pydantic-ai :)