Description

Currently, the OpenAIError::StreamErrorvariant only carries a String payload.

This makes it hard to match the error, and when I log the error (even with '{:?}') it is hard to get a meaningful information for debugging.

Problem

If I send an invalid request to the server and receive an error. The response body often holds information about why the request was invalid. For example, if I set both max_tokens and max_completion_tokens, I get the error:

{
  "error": {
    "message": "Setting 'max_tokens' and 'max_completion_tokens' at the same time is not supported.",
    "type": "invalid_request_error",
    "param": "max_tokens",
    "code": "invalid_parameter_combination"
  }
}

But when I print the message, I get stream failed: Invalid status code: 400 Bad Request.

The only way to get the original error message is to use the verbose reqwest::Client and setting the log level for reqwest at trace. However, this is obviously not a good solution.

Proposed Solution

Right now, I could only find three places where this error variant is initialized:

Twice in Client::stream and Client::stream_mapped_raw_events, where the inner reqwest_eventsource::Error is converted into a String via to_string() call.

And once in AssistantEventStream::try_from, where an unrecognized event (eventsource_stream::Event) is pretty-printed into a String

Therefore, it should be relative to define a new enum:

pub enum StreamErrorKind {
    #[error("{0}")]
    ReqwestEventSource(reqwest_eventsource::Error),
    #[error("Unrecognized event: {0:#?}")]
    UnrecognizedEvent(eventsource_stream::Event),
}

And change the StreamError payload to this type.