fix: count tokens from tool definitions when adjusting for context window #942
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…ndow Signed-off-by: Grant Linville <grant@acorn.io>
g-linville
changed the title
g-linville
requested review from
StrongMonkey,
drpebcak,
iwilltry42,
njhale and
thedadams
Member
Author
|
Moved to draft. I'm going to use a proper tokenizer library so that we can get accurate token counts. |
Member
Author
|
There is more work that would be needed here and it's just not a priority right now. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
The tokens used for tool definitions are counted as part of the context window, but we were failing to count them here when trying to stay beneath the limit. This introduces a fix so that we account for it.
Our token estimation still isn't great. For example, a message consisting of
1\nrepeated 30000 times is counted as 20000 tokens in our method (since we just divide the character count by 3), but in OpenAI's tokenizer, it is 60000 tokens, meaning we grossly underestimate. I think we should consider calling out to Python to use thetiktokenlibrary to count tokens for us, as that will be far more reliable. However, in practice, this rarely seems to cause us any trouble, so maybe we are fine sticking with our current approach of just dividing by 3.