Do not expose providers api

[brichet]

This PR avoids exposing the chat or LLM model API, to avoid exposing API keys or other secrets that are exposed in langchain model (e.g. https://github.com/langchain-ai/langchainjs/blob/1846bdbb0fbe88c0981d11bfc5e302d2756565f2/libs/langchain-mistralai/src/chat_models.ts#L911).

The chat model only expose a stream() wrapper, and the completer model expose a fetch() wrapper.

cc. @trungleduc who first catches this issue and suggest a workaround.

[trungleduc]

Neat!

[trungleduc]

And you are tagging the wrong person 😸

[brichet]

And you are tagging the wrong person 😸

Oups

[jtpio]

to avoid exposing API keys or other secrets that are exposed in langchain model

That wouldn't prevent a plugin to access _completer directly in JavaScript, or with provider['_completer']?

[trungleduc]

to avoid exposing API keys or other secrets that are exposed in langchain model

That wouldn't prevent a plugin to access _completer directly in JavaScript, or with provider['_completer']?

you don't have access to the provider

[brichet]

to avoid exposing API keys or other secrets that are exposed in langchain model

That wouldn't prevent a plugin to access _completer directly in JavaScript, or with provider['_completer']?

The _currentProvider exposed in AIProviderRegistry is only the description of the provider, but the models are not instantiated.
After this PR, the currentCompleter and currentChatModel are only objects wrapping the required functions.

[brichet]

Actually I wonder if it won't break the WebLLM provider, since some of its methods may not be available anymore.

[brichet]

@trungleduc @jtpio there is now a flag to allow exposing the whole model API when registering the provider.
This flag is used for WebLLM, which needs some custom API functions (there is no sensitive information available in its API)

ai/src/default-providers/index.ts

Line 157 in d075ec5

void model.initialize(report => {

To prevents being able to modify this flag, the AIProviders, _currentProvider are not exposed anymore in the the registry.
The _name attribute has also been moved in a private namespace, to prevent inconsistency between the name and the models. It could also lead to expose the chat model while not expected.

[jtpio]

Thanks @brichet! Yes exposing less things sounds good.