Research Assistant uploaded CSV files "are non-existent"

jimbo

I upgraded myself to the ultimate tier to test the python interpreter with my data.

I have CSV file which is 1.1MB and contains data, sample of which is below. When I upload it the research assistant cannot access it and I see: Sources provided /user_attached_documents/prices(1).csv are non-existent files. I do not see an error during the upload. And the size limit is well below the limit of 16MB stated in the documentation.

The file upload works with random pdf and csv files I found on the internet.

prices(1).csv (full year of data, 15 minute intervals), only few first lines included:

hour,price
2025-01-01T00:00:00.000Z,0.496
2025-01-01T00:15:00.000Z,0.496
2025-01-01T00:30:00.000Z,0.496
...

Since the documentation is my source of truth and it states that CSV file uploads under 16MB are supported, I'd expect it to work. In cases of error, I would expect to see error message either in the file uploader or the assistant should tell me what's causing the error so I can avoid it in the future.

While trying to investigate, I found this thread which is possibly related: https://kagifeedback.org/d/2813-allow-upload-to-more-than-assistant If I understood that thread correctly, there may be other limits than just the file size limit.

jimbo

I did some testing by copying more and more lines to test[index].csv and asked whether the assistant can see it. It could see test2.csv but not test3.csv but it also could not suggest/think why that could be the case. So the file upload works, but with too big files it fails without errors or messages. It seems that I wrongly assumed that it would work with any file under 16MB.

yiwei-1

Hi, just to clarify: right now, the Python interpreter doesn’t have direct access to the files you upload, and it can’t export a CSV file for you to download (though this is planned).

We do plan to enable the assistant to use the Python interpreter to manipulate and update user-uploaded files. Even a relatively small CSV like this can easily contain _665K tokens (at 19 tokens per line), which already exceeds the input limits of many models if we pass the entire file directly.

For now, I recommend asking the assistant to write a Python script based on your data schema which you can then run locally.