LiteLLM support

[JoeSaelens]

Description of feature

I’m trying to use CellAnnotator with LiteLLM as the model provider, but I couldn’t find a way to set an OpenAI-compatible base_url for a LiteLLM proxy.

Are there plans to support LiteLLM? If so, what configuration would you recommend?

[airouter-dev]

I traced the current provider path, and this is not configurable today: OpenAIProvider initializes _base_url to None, while get_provider() only accepts an API key. The OpenAI SDK wiring already passes base_url when that internal field is set, so a separate LiteLLMProvider should not be necessary.

A provider-neutral shape would be:

CellAnnotator(
    ...,
    provider="openai",
    model="cell-annotation-model",  # a LiteLLM model name/alias
    api_key="...",
    base_url="http://localhost:4000/v1",
)

Implementation details that seem important:

1. Add base_url: str | None to OpenAIProvider, get_provider(), LLMInterface, and the annotator constructors, then propagate it to every child SampleAnnotator.
2. Do not reuse the cached openai provider when a custom URL is supplied; endpoint-specific providers need separate client instances.
3. Add a test asserting that the OpenAI client receives the custom URL, plus a propagation test for child annotators.
4. Account for LiteLLM aliases in model discovery. The current OpenAI filter only keeps IDs containing gpt or o1, so a valid alias such as cell-annotation-model would disappear from list_available_models(). For custom OpenAI-compatible endpoints, filtering only known non-chat categories (embeddings, TTS, etc.) is safer.

Relevant current code: OpenAI client construction, model filtering, and provider factory/cache.

Disclosure: I work on AI Router, another OpenAI-compatible endpoint, so I reviewed this from the gateway integration side. The suggested interface is intentionally endpoint-neutral and would cover LiteLLM and other compatible services without adding vendor-specific branches.