When testing the Gaia dataset, the issue of tokens exceeding the limit often occurs.

[Yuzt-zju]

This model's maximum context length is 128000 tokens. However, your messages resulted in 229568 tokens. Please reduce the length of the messages.

[ryantzr1]

hi @Yuzt-zju, I think this has something to do with the context that is passed around. will have to do a deep dive on this.

[ryantzr1]

update: currently testing on the GAIA benchmark so probably will do a comprehensive review of the codebase soon.

[Yuzt-zju]

update: currently testing on the GAIA benchmark so probably will do a comprehensive review of the codebase soon.

In fact，this issue occurs more often which makes many tool calls fail:
Failed: count_attendees() takes from 0 to 1 positional arguments but 3 were given\ud83d\udd0d Executing: calculate_potato_consumption \u2192 Success: 2...\ud83d\udd0d Executing: convert_to_bags \u2192 Failed: convert_to_bags() takes from 0 to 1 positional arguments but 3 were given

[ryantzr1]

@Yuzt-zju do you happen to have a solution to the token limit issue? I'll try to fix it this week.

[Yuzt-zju]

@ryantzr1 My token limit might be due to restrictions imposed by the service provider, and it may be possible to lift the limit by making a request for this. I'm not sure if you are encountering the same issue. The cause is likely the excessively long context that was added, and I don't have a good solution for this.

[ryantzr1]

hmm I pushed a fix here via branch: fix/content-length-exceed-window

I think there are two main issues that we are facing now:

1. browser use is not good enough
2. Tool creation needs to improve too

will try to fix them