When I do the manual drag-and-drop of a file into the prompt to upload to my assistant during a session, it is REMARKABLE how well it can analyze the data set
When I try to upload the same size file programmatically through a function call mid session, it is remarkable at how terrible it is.
I can’t seem to find any documentation to explain why the experiences couldn’t be further apart. This basically makes any dynamic data upload during a session DOA
My guess is “Time” when you drag and drop you give the chunking and data embedding system time to process the file, when you programmatically upload it in real-time, you do not.
Another angle would be that a manually attached file is added to the context window of the assistant in contrast to a file that has been added programmatically where the model reads only the first 1000 characters.
@brdemorin do you have Code Interpreter and retrieval enabled? Which model are you using?