Gpt-4-1106-preview does not read the whole file like web ChatGpt4 does

Hi there,

I am working with pdfs, on the web version I was able to give it a pdf and ask for a synopsis and it based its synopsis from the entire document. While when I use the API it clearly only basis its response of less than the first 10% of the file, based of its response sometimes containing “Based on the first 6% of the file”.

Am I misusing it? Am I missing something special that I need to do before it will consider using the whole file?

My prompt clearly states that its response needs to be based off the entire file, that it should read the whole file prior to making a response.

What can I do to make it use the whole file like the web version?

Thanks in advance.

My guess is that you have Code Interpreter turned on and it is that that is looking at the first 50 or so lines of your pdf by converting it to text. If you disable CI and upload your files to the retrieval system then it should function as expected:

I only have tools=[{"type": "retrieval"}], unfortunately.

How big is the file?

(extra words for the minimums bot)

~4mb, ~30k words, should fit within the 2 million tokens for files right?

I have the same problem, I have only retrieval turned on but when I upload a file it doesnt read it all and continuously asks me if he can answer my question based on what it did read so far or if it should continue to read the document.
It even writes the percentage of the document read so far.

1 Like