Add component config support

[jackgerrits]

Component config adds a generic way for classes to opt in to being declaratively specified.

For example:

from __future__ import annotations

from autogen_core import Component
from pydantic import BaseModel


class Config(BaseModel):
    value: str


class MyComponent(Component[Config]):
    component_type = "custom"
    config_schema = Config

    def __init__(self, value: str):
        self.value = value

    def _to_config(self) -> Config:
        return Config(value=self.value)

    @classmethod
    def _from_config(cls, config: Config) -> MyComponent:
        return cls(value=config.value)

An example component config json:

{
    "provider": "autogen_ext.models.openai._openai_client.AzureOpenAIChatCompletionClient",
    "component_type": "model",
    "version": 1,
    "description": null,
    "config": {
        "values": {
            "azure_deployment": "{your-azure-deployment}",
            "model": "gpt-4o",
            "api_version": "2024-06-01",
            "azure_endpoint": "https://{your-custom-endpoint}.openai.azure.com/",
            "azure_ad_token_provider": {
                "provider": "autogen_ext.models.openai._azure_token_provider.AzureTokenProvider",
                "component_type": "token_provider",
                "version": 1,
                "description": null,
                "config": {
                    "provider_kind": "DefaultAzureCredential",
                    "scopes": [
                        "https://cognitiveservices.azure.com/.default"
                    ]
                }
            }
        }
    }
}

Which is a dump of:

az_model_client = AzureOpenAIChatCompletionClient(
    azure_deployment="{your-azure-deployment}",
    model="gpt-4o",
    api_version="2024-06-01",
    azure_endpoint="https://{your-custom-endpoint}.openai.azure.com/",
    azure_ad_token_provider=AzureTokenProvider(DefaultAzureCredential(), "https://cognitiveservices.azure.com/.default"),
)

[ekzhu]

    def _to_config(self) -> Config:
        return Config(value=self.value)

    @classmethod
    def _from_config(cls, config: Config) -> MyComponent:
        return cls(value=config.value)

This returns a Pydantic basemodel, but as we discussed earlier Component types might be an implementation detail.

@jackgerrits @victordibia let's discuss this so we have a coherent way across configs and states.

[jackgerrits]

    def _to_config(self) -> Config:
        return Config(value=self.value)

    @classmethod
    def _from_config(cls, config: Config) -> MyComponent:
        return cls(value=config.value)

This returns a Pydantic basemodel, but as we discussed earlier Component types might be an implementation detail.

@jackgerrits @victordibia let's discuss this so we have a coherent way across configs and states.

_to_config returns a type with a lower bound of base model (as specified by the Component[T] generic. The value passed to the _from_config function is guaranteed to be a valid instance of that type by the framework.

These are only ever called as part of the framework so it is safe to use full types here

[jackgerrits]

I think the distinction between config and state is an important one. I added a section in the docs about it with my take but I think we should talk more about it!

How does this differ from state?

This is a very important point to clarify. When we talk about serializing an object, we must include all data that makes that object itself. Including things like message history etc. When deserializing from serialized state, you must get back the exact same object. This is not the case with component configuration.

Component configuration should be thought of as the blueprint for an object, and can be stamped out many times to create many instances of the same configured object.

[ekzhu]

Created a separate PR for the ChatCompletionClient argument list: #4774

[victordibia]

@jackgerrits
Pydantic is generally unhappy with any fields that start with model_ such as model_info, model_capabilties, I get a flurry of warnings ...

https://github.com/microsoft/autogen/blob/7b02d9dbf5c9e10ed3f2062eb09732b60ed44eec/python/packages/autogen-ext/src/autogen_ext/models/openai/config/__init__.py#L35

Some solutions

• avoid using any field with model_ (doesnt feel quite right as it will interfer with core behaviours especially in component loading)
• Use the "hack here" ... , add model_config = ConfigDict(protected_namespaces=()) to the data class

[jackgerrits]

Pydantic loosened their requirements of naming in 2.10, so we now require >=2.10 for pydantic to avoid these warnings. See this PR: #4811

The warnings all go away with that version, you might need to uv sync your environment

[victordibia]

Ah .. got it!
I just need to upgrade my pydantic version!