Inconsistent PDF handling in The Assistant

fs1010

PDF uploads seem to be treated differently from links to PDFs. I recently tried uploading the “Attention is all you need” PDF and ran into an issue with it hitting the maximum input length, despite it being only around 10000 tokens/15 pages. This issue disappears when providing the Arxiv link to the PDF instead of uploading it yourself. I made sure it extracted the full content in both examples.

https://kagi.com/assistant/f5a8b9b7-3e35-4197-91f9-2845e4c33ff7

https://kagi.com/assistant/e6f2481c-a7b0-41f7-8d92-74702f0d5a86

https://kagi.com/assistant/ad57e3c3-20a2-4a4b-93dc-f8dd9cf874d9

I expect uploaded PDFs to be processed in the same way as links to PDFs (aka for it to actually work).

yiwei-1

Could you check whether the PDF file is valid and not corrupted? Its MIME type should be recognized as application/pdf rather than application/octet-stream.

fs1010

Turns out this may need to go in Orionfeedback instead. It seems PDF saving is not working correctly in Orion. The PDF is saved correctly in safari (has a proper thumbnail) while Orion corrupts it (question mark icon). The PDF saved from safari is the only one that works.

https://kagi.com/assistant/18083d0a-955a-4633-bf69-46c2aac24f0f