Hacked in support for azure. Have only tried the chat completion endpoint.

Limitations:

• One client per deployment (e.g. no way to pass in deployment on a per-invocation) NOTE: currently accepting deployment_id as model on per-method basis to address this limitation
• Typing busted for constructor (because of the use of *args, **kwargs) - can be fixed by copying the signature from the base (or possibly using a functools.wraps like decorator that copies the signature)
• No typing support for azure specific additions

Additional work:

• Audio/whisper does not work for openai or azure currently. Both services expect the filename provided in the second case, but the v1 API doesn't have a parameter for that
  • send io.BufferedReader: RuntimeError: Expected entry at file to be bytes or a tuple but received <class '_io.BufferedReader'> instead.
  • send bytes: openai.BadRequestError: Error code: 400 - {'error': {'message': "Unrecognized file format. Supported formats: ['flac', 'm4a', 'mp3', 'mp4', 'mpeg', 'mpga', 'oga', 'ogg', 'wav', 'webm']", 'type': 'invalid_request_error', 'param': None, 'code': None}} files = [('file', ('hello.m4a', files[0][1], 'application/octet-stream'))]
• Lacks good exception support, especially for prompts caught by the content filter (WIP consider a custom exception for rai #14)
• No Module client support
• No CLI support
• Some response types still need wrapping to expose Azure-specific properties
  • Model
  • ??
• No testing for fine-tuning/fine-tunes/files yet
• Typing improvements, where possible. Needs a thorough type checking/linting/blackening of the code.
• Need blessings for where we're touching internal APIs, where we're conflating deployment_id with model
• General clean-up of code, address TODOs