Multimodal (Vision) support

[liwii]

Now some OpenAI LLMs support image inputs, so it'd be great if we could support evaluation with image inputs.

Ref: https://platform.openai.com/docs/guides/vision

The goal is to update EvalClient interface and allow metrics with image inputs like the following:

prompts = ["What is in the image?", "What is in the image?",...]
generated_outputs = ["A green parrot flying away from...", "A huge robot woking for...", ...]
images = [
  # Link to the image or base64 encoded image
  "https://...",
  "data:image/jpeg;base64,b656...",
  ...
]

results = langcheck.metrics.answer_relevance_with_input_image(
  generated_outputs=generated_outputs,
  prompts=prompts,
  image_urls=image_urls,
  eval_model=eval_client
)

Probably there are more things to discuss before actually shipping this feature, but we can prototype and run some metrics first.

[liwii]

@matchcase

This is a bit abstract, but it would be also interesting to work on this!!