Describe the Bug
I previously used the Perplexity API to build a script that parses multiple PDFs to extract specific information. When I first tested it in August, everything worked correctly and the API was able to read the PDFs without issues.
After modifying my prompt and rerunning the script, the API stopped reading the PDFs and/or started hallucinating responses. I’ve tested several variations of the example code from the docs (https://docs.perplexity.ai/guides/pdf-uploads). My testing code is attached (with the key and links removed).
Originally, I used Dropbox direct download links, which worked fine at first but no longer do. I also tried Google Drive, S3, and GitHub Pages links, but none of them worked.
Expected Behavior
Return a summary of the attached pdf.
Actual Behavior
API Output:
Extraction: I’m unable to access or view the contents of the attached PDF (“link”), so I cannot extract or summarize any verbatim data from it.
Steps to Reproduce
-
Call the API with the following request:
STRICT EXTRACTION TASK:
You must ONLY extract and return data that is a direct, verbatim copy of text or tables in the ATTACHED PDF.
FORBIDDEN:
- No inference, suggestions, guesses, or external data.
- Do NOT output anything not fully, explicitly written in the PDF.
- If a value is missing, ambiguous, or not clearly labeled, return “”.
TASK:
Summarize the document.
-
Test with a dropbox direct download link.
-
Observe the unexpected behavior.
Environment
- API Model: sonar-pro
- Operating System: Tested on both MacOS and Windows]