JSONDecodeError - OpenAI API response is not json compliant

[jpabbuehl]

Hello, many thanks for the library

OpenAI sometimes returns response that is not json compliant

• Adding trailing comma
• Double quoting without backslashes
• Not closing bracket such xxx."\n}'

Tested with gpt-3.5-turbo-0613

When using pydantic model as output_schema (inspired from the example notebook), orjson.loads fails. Unfortunately, it's a bit unpredictable and cannot write a reproducible example. Any advice?

Many thanks

[pors]

It would be cool to add a mechanism like this from langchain: https://python.langchain.com/docs/modules/model_io/output_parsers/output_fixing_parser

[jorahn]

i'm encountering the same or a similar error, but i'm not using pydantic output schema, just plain prompts and a system prompt (programming related, the result includes correctly rendered python markdown). i believe this is the relevant part of the stack-trace:

    for chunk in ai.stream(message, system=system_prompt):
  File "/home/user/.local/lib/python3.10/site-packages/simpleaichat/chatgpt.py", line 160, in stream
    chunk_dict = orjson.loads(chunk)
orjson.JSONDecodeError: Input is a zero-length, empty document: line 1 column 1 (char 0)

zero-length input should be caught by https://github.com/minimaxir/simpleaichat/blob/55b8324f8f4558c42e33ab6478446af6fd1773d7/simpleaichat/chatgpt.py#L157 though...

[agentzh]

I ran into this today, too. It seems like the library fails to handle errors returned by the API (like a 429 Too Many Requests response). It definitely needs more robust error handling and error reporting... Otherwise, it's very confusing.