fix: openai per batch limit #730
Merged
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
a736ac4 to
9567c68
Compare
7fda62f to
5759fc8
Compare
5759fc8 to
a4ae689
Compare
a4ae689 to
2293e50
Compare
2293e50 to
564b980
Compare
564b980 to
d15c3a6
Compare
d15c3a6 to
ab5a5c6
Compare
ab5a5c6 to
1a47a33
Compare
1a47a33 to
edc03c6
Compare
edc03c6 to
09a1437
Compare
Comment on lines
+153
to
+164
| def _estimate_token_length(self, document: str) -> float: | ||
| """ | ||
| Estimates token count based on UTF-8 byte length. | ||
| """ | ||
|
|
||
| total_estimated_tokens = 0 | ||
| for char in document: | ||
| byte_length = len(char.encode("utf-8")) | ||
| total_estimated_tokens += byte_length * 0.25 # 0.25 tokens per byte | ||
|
|
||
| return total_estimated_tokens | ||
|
|
There was a problem hiding this comment.
I would suggest that we do this, and then remove the type changes that this propagates:
Suggested change
| def _estimate_token_length(self, document: str) -> float: | |
| """ | |
| Estimates token count based on UTF-8 byte length. | |
| """ | |
| total_estimated_tokens = 0 | |
| for char in document: | |
| byte_length = len(char.encode("utf-8")) | |
| total_estimated_tokens += byte_length * 0.25 # 0.25 tokens per byte | |
| return total_estimated_tokens | |
| def _estimate_token_length(self, document: str) -> int: | |
| """ | |
| Estimates token count based on UTF-8 byte length. | |
| """ | |
| total_estimated_tokens = 0 | |
| for char in document: | |
| byte_length = len(char.encode("utf-8")) | |
| total_estimated_tokens += byte_length * 0.25 # 0.25 tokens per byte | |
| return ceil(total_estimated_tokens) |
There was a problem hiding this comment.
I don't think that works because that would mean it rounds each individual documents tokens not the whole batch. The api however counts the bytes of everything first and then rounds.
E.g. if you send 100 batches of just "a" it would be estimated with 100 tokens this way but the api only assigns 25 (which makes no sense but well that's how it works)
There was a problem hiding this comment.
Hmmm... interesting. I didn't realise that on first inspection.
JamesGuthrie
approved these changes
May 14, 2025
|
@Askir Great catch! How did you find out about OpenAI's token estimator? |
|
@Michnic120 I just spent a full day trying a bunch of tokenized inputs to figure out what works and what doesn't. If you break the limit, the error message contains the token count that their api is "calculating" so eventually I figured out how it worked simply through trial and error. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Fixes #728
The magic here is that openai has a token estimator in front of their actual API this one simply counts utf-8 bytes and for each one calculates 0.25 tokens.
This is what the 300k limit per batch request limit applies to. The actual tokens don't matter. A request with 1 million tokens and 1.2 million bytes will go through just fine as the estimator thinks this is exactly 300k "tokens".
Likewise a request with 200 000 tokens but 1.5 million bytes will fail.