Question About GPT_Structured_Request and IndexError Bug

[Cyan9644]

Hi @chowington ,

I noticed that there is no call to the update_cost method (from openai_logger_singleton) in the GPT_Structured_Request function (persona/prompt_template/gpt_structure); this uses the same framework as ChatGPT_Structured_Request, which does have the call to the logger. This function is used indirectly in multiple places, including with long prompts like generate_focal_pt_v1. Not sure if I'm this is unintended or I'm just missing something, but I thought I'd ask. On a related note, I found that there is a bug in safe_generate_structured_response where on long (350+ step) runs, the agents will sometimes mistakenly generate incomplete plans. I don't know why this happens, but it is easily fixed with regex as shown below in case anyone else has experienced this too. Thanks!

if(isinstance(response_format, NewSchedule)):
delimiters = ["pm", "am", "PM", "Pm", "P.M", "p.m", "P.m", "AM", "Am", "A.M", "A.m", "A.m"]
regex = map(re.escape, delimiters)
regex_pattern = '|'.join(regex)
splitter = re.split(regex_pattern, str(curr_gpt_response))
if(endtime is not None):
if not(endtime in (splitter[-1])):

        ...

       ( i-=1
        time.sleep(1)
        continue)

Edit: Cost Logger is correct as is according to API statistics

Cyan