A.I. Hangs consistently Between 30 and 40 minutes of chat

Not sure why, but the chatbot always consistently hangs at 32 to 40 minutes into the conversation. When it hangs, it gets into a routine where it delivers repeatedly the previous response. I can only end this repetition loop by re-booting.

Regarding token length, I have experimented with topic, temperature, and token length (nearly eliminating token length to next to nothing) That said, check out these usage stats. In spite of my keeping prompts to a minimum, the token length in prompts seems excessive. If 1,000 tokens is about 750 words and a paragraph of 34 words = 35 tokens there is absolutely no way I submitted a prompt of 18,000+ tokens and 23K tokens and 10K tokens. I would be impossible to even carry a conversation in such a manner.

5:25 AM
gpt-3.5-turbo-0301, 7 requests
18,657 prompt + 290 completion = 18,947 tokens

5:30 AM
gpt-3.5-turbo-0301, 7 requests
23,058 prompt + 327 completion = 23,385 tokens

5:35 AM
gpt-3.5-turbo-0301, 3 requests
10,964 prompt + 101 completion = 11,065 tokens.

As a test, I just sent a message “Hi Chatty, this is a test prompt.” The response was "Hello there! How can I assist you today?

gpt-3.5-turbo-0301, 1 request
200 prompt + 10 completion = 210 tokens

This seems normal – what could have been happening?


Notice it says 7 requests here… it seems like the big ones are where you append old messages maybe? What’s your cut-off point for that?

This is a plugin that I purchased off Unreal Engine marketplace. I suspected it appended old messages, which gives the chatbot the semblance of memory.

I do not know what the cut-off is for this. But it seems like capping the chat might solve my problem when I hit the rate limit ceiling (assuming that is causing the glitch), it would have deleterious effects on the chat experience.

But you’re right–that cut-off is likely in the plugin’s source code somewhere. I didn’t see an adjustable cut-off variable in the dash panel. That said, once I locate it, I wonder what the sweet spot would be.