JSON Parser streaming seems to fail depending on model used about opengpts HOT 1 OPEN

Hi

I found that the JsonOutputFunctionsParser worked with the latest OpenAI 3.5 chat model in streaming, but not with the Azure equivalent. I found the problem was that the "first" item sent in a stream for the function call arguments had a value of '' (empty string) for the OpenAI model, and a value of None for the Azure equivalent. The empty string is fine when being converted to json, but the value of None can not be converted to json by json.loads. So the parser throws an error with the Azure OpenAI version.

I'm not sure what the root cause is - perhaps something in the way the streaming output is generated for the Azure OpenAI versions? As a workaround in case that's helpful for others, I created an output parser which takes the JsonOutputFunctionsParser class as its base and overrides the parse_result function to handle the problem. The file is in a Github repo - the function is exactly the same as for JsonOutputFunctionsParser except for the addition of the following lines (24 - 26):

# Convert None value to string to allow conversion to JSON.
if function_call["arguments"] is None:
    function_call["arguments"] = ''

from opengpts.