tiktoken_rs is a bit behind (and even upstream tiktoken doesn't have all
of these models)
We were incorrectly using the cl100k tokenizer for some models that
actually use the o200k tokenizers. So that is updated.
I also made the match arms specific so that we do a better job of
catching whether or not tiktoken-rs accurately supports new models we
add in.
I will also do a PR upstream to see if we can move some of this logic
back out if tiktoken better supports the newer models.
Release Notes:
- Improved tokenizer support for openai models.
Ben Brandt
3a3d3c05e8