Is it planed to increase the number of files which can be uploaded per Assistant in the Assistant-API.
The 512MB per file is not necessary but 20 files is way to less.
Are you actually hitting such a limit?
Vector stores don’t have an intrinsic low file limit like ChatGPT. Each can have 10000 files, files of 5M extracted tokens.
Maybe it is just the “attach” if you are putting them in a thread message?
this is the limit for code_interpreter, not for file_search …
do you really need more than 20 files for code interpreter purposes ? I’m afraid for now you don’t have any other choice than handle the list of files used by code interpreter (removing files you don’t need anymore for instance)
I believe there may be confusion here around these limits. If I understand it correctly, when using file_search tool, the system imposes a limit of 16k of context to be used for file_search results. Since each chunk returned is typically 800 bytes, this implies a limit of 20 chunks (not files).
That would be the “bugdet”
The
file_searchtool currently has a token bugdet of:
- 4,000 tokens for
gpt-3.5-turbo- 16,000 tokens for
gpt-4*models
This is only perceptible in the quality of the answers, though, not a file input limit.
Some of the confusion may derived from the fact that there is a setting on the file_search tool that limits the number of results. You can see that manifested in the Assistant playground UI as shown in this screenshot.
I understand that this imposes an additional limit on the number of results – presumably as a way to allow you to limit the number of input tokens used for those results.
My point was simply that there are various limits around that can be easily misinterpreted.
