openai[patch]: instantiate clients lazily

[ccurme]

#29925

Simplified version of the change below.

Before:

from typing import Any, Optional

import openai
from pydantic import BaseModel, Field, SecretStr, model_validator
from typing_extensions import Self


class BaseChatOpenAI(BaseModel):
    client: Any = Field(default=None, exclude=True)
    async_client: Any = Field(default=None, exclude=True)
    my_special_api_key: Optional[SecretStr] = Field(alias="api_key")

    @model_validator(mode="after")
    def validate_environment(self) -> Self:
        client_params = {"api_key": self.my_special_api_key}
        if not self.client:
            self.client = openai.OpenAI(**client_params)  # type: ignore[arg-type]
        if not self.async_client:
            self.async_client = openai.AsyncOpenAI(**client_params)  # type: ignore[arg-type]
        return self

After:

from typing import Any, Dict, Optional

import openai
from pydantic import BaseModel, Field, PrivateAttr, SecretStr, model_validator
from typing_extensions import Self

class BaseChatOpenAI(BaseModel):
    _client: Any = PrivateAttr(default=None)
    _async_client: Any = PrivateAttr(default=None)
    _client_params: Dict[str, Any] = PrivateAttr(default_factory=dict)
    my_special_api_key: Optional[SecretStr] = Field(alias="api_key")

    # Needed to support initializing the class with `client`
    def __init__(
        self, client: Any = None, async_client: Any = None, **kwargs: Any
    ) -> None:
        super().__init__(**kwargs)
        self._client = client
        self._async_client = async_client

    @model_validator(mode="after")
    def validate_environment(self) -> Self:
        self._client_params = {"api_key": self.my_special_api_key}
        return self

    @property
    def client(self) -> Any:
        if self._client is None:
            self._client = openai.OpenAI(**self._client_params)  # type: ignore[arg-type]
        return self._client

    @client.setter
    def client(self, value: Any) -> None:
        self._client = value

    @property
    def async_client(self) -> Any:
        if self._async_client is None:
            self._async_client = openai.AsyncOpenAI(**self._client_params)  # type: ignore[arg-type]
        return self._async_client

    @async_client.setter
    def async_client(self, value: Any) -> None:
        self._async_client = value

I'm not sure there is a way to do this that is not semver-breaking. One usage pattern I know breaks is subclasses of BaseChatOpenAI, most of which have unique names for their API keys and other attributes (e.g., xai_api_key, together_api_key, etc.). These subclasses have post init steps that look like:

@model_validator(mode="after")
def validate_environment(self) -> Self:
    client_params = {"api_key": self.xai_api_key.get_secret_value()}
    if not self.client:  # <-- breaks
        self.client = openai.OpenAI(**client_params)

The check if not self.client breaks because we now try to instantiate the client in a property, and this requires that client params be accessible to the property.

Current plan is to update classes we control and explain how to update sub-classes in release notes. The update is simple but would be annoying.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
langchain	✅ Ready (Inspect)	Visit Preview	💬 Add feedback	Feb 22, 2025 3:32pm

[baskaryan]

looks reasonable to me, assuming we re-ran the benchmarks and this fixed the issue?

[baskaryan]

ooc are these untyped on purpose?