I slightly modified parse_asterisks_brackets to pass asterisks - because I had the same issue with asterisks, because the LLM was charging me for generations that base_llm.py was throwing in the bin, resulting in conversation loops of
Me: Actual conversation attempt.
LLM; *responds with perfectly valid RP Action*
Pantella: (Thrown in the bin by Line 162.)
Waiting for player response....
Me: ...
LLM; *responds with the same perfectly valid RP Action*
Pantella: (Thrown in the bin by Line 162.)
Waiting for player response....
Me: *waits while you continue to think.*
LLM; *responds with a slight variation on a perfectly valid RP Action*
Pantella: (Thrown in the bin by Line 162.)
Waiting for player response....
Once I tweaked parse_astrersks_brackets to NOT return
sentence = ''
by commenting that out
The llm would sometimes use . then some text then . again.
So I modified parse_asterisk_brackets to catch and sub for .
That resulted in:
2024-04-04 19:57:32 INFO| next_author detected as: . Laura
2024-04-04 19:57:32 INFO| Next author is not a real character: . Laura
2024-04-04 19:57:32 INFO| Retrying...
2024-04-04 19:57:32 ERROR| LLM API Error: Invalid author
2024-04-04 19:57:32 INFO| Retrying without saying error voice line
Over and over and over
Line 347 appears like it increments retries by one BUT
Then the IF/ELIF/ELSE comes RIGHT BACK HERE and increments retries by 1 again.
It doesn't appear to ever reach any instance of an error
The IDEAL situation is that if you get into a position where there is no valid author like this Catch the 5th retry and
Send the conversation for summarization
I lost about 30 minutes of really excellent, amazing conversation for summarization because of an invalid speaker.
And the LLM went through eighteen retries before I shut it off manually with the Launcher...but if I'd stepped away, it could have easily drained my account, $0.00251 cents at a time. 😹
base_llm.py -> base_llm -> clean_sentence -> parse_asterisks_brackets
I slightly modified parse_asterisks_brackets to pass asterisks - because I had the same issue with asterisks, because the LLM was charging me for generations that base_llm.py was throwing in the bin, resulting in conversation loops of
Me: Actual conversation attempt. LLM; *responds with perfectly valid RP Action* Pantella: (Thrown in the bin by Line 162.) Waiting for player response.... Me: ... LLM; *responds with the same perfectly valid RP Action* Pantella: (Thrown in the bin by Line 162.) Waiting for player response.... Me: *waits while you continue to think.* LLM; *responds with a slight variation on a perfectly valid RP Action* Pantella: (Thrown in the bin by Line 162.) Waiting for player response....
Once I tweaked parse_astrersks_brackets to NOT return sentence = '' by commenting that out The llm would sometimes use . then some text then . again.
So I modified parse_asterisk_brackets to catch and sub for .
That resulted in:
2024-04-04 19:57:32 INFO| next_author detected as: . Laura 2024-04-04 19:57:32 INFO| Next author is not a real character: . Laura 2024-04-04 19:57:32 INFO| Retrying... 2024-04-04 19:57:32 ERROR| LLM API Error: Invalid author 2024-04-04 19:57:32 INFO| Retrying without saying error voice line
Over and over and over
Line 347 appears like it increments retries by one BUT Then the IF/ELIF/ELSE comes RIGHT BACK HERE and increments retries by 1 again. It doesn't appear to ever reach any instance of an error
The IDEAL situation is that if you get into a position where there is no valid author like this Catch the 5th retry and Send the conversation for summarization
I lost about 30 minutes of really excellent, amazing conversation for summarization because of an invalid speaker.
And the LLM went through eighteen retries before I shut it off manually with the Launcher...but if I'd stepped away, it could have easily drained my account, $0.00251 cents at a time. 😹