I am using an API assistant created in Playground. I uploaded a JSON file containing a list of references with various attributes. When I query the assistant about the number of references in its database (JSON), it returns a number that is usually below 10, despite the actual count being over 1000. The code interpreter option is enabled. When I performed the same test or a similar one in ChatGPT, I uploaded the file and asked the same question, and the result was highly accurate. Why is the API assistant not providing the correct results?
It’s hard to say. We don’t know that much about the retrieval framework.
What I can say though is that the ChatGPT version most likely has a more up-to-date framework while the Assistant’s is most likely outdated/sitting at a stable checkpoint.
I’m confused by this though. Are you uploading this JSON file directly or storing it for retrieval?
If you are uploading it and asking the model to count it using Code Interpreter the results should be the exact same. Lastly, you should be able to see the code that it used and can easily compare.
1 Like