Open gautierdag opened 11 months ago
No, this is fair.
The additional constraints are a) logit_bias
does have a limit on the number of tokens you can specify (I think 100) and b) appropriate tool accuracy will be bad when you have a ton of tools so it's mostly pretecting the user from themselves. Even OpenAI Plugins doesn't let you use more than 3.
However, with gpt-3.5-turbo-0613
the latter assumption may not be valid, so it may be worth experimenting.
Hi, Found this library through your post, and I just wanted to say a great job at hitting it on the nail!
Reading the docs (https://github.com/minimaxir/simpleaichat/blob/db19d26250cdee7db8acb52027632083e83aeb28/PROMPTS.md?plain=1#L59C1-L59C185):
I couldn't help but point out that the tokenization scheme in
gpt-3.5-turbo
encodes the first 999 numbers as single tokens. Therefore, up to 999 tools should be possible using"max_tokens":1
. It would just mean that you'd need to deal differently with the logit bias since the tokens for n > 9 are indeed not sequential. So the trick below wouldn't work:https://github.com/minimaxir/simpleaichat/blob/db19d26250cdee7db8acb52027632083e83aeb28/simpleaichat/chatgpt.py#L189)
Feel free to close this.