I have been doing some work in ChatBot API AI sites regarding reading a pdf file. I have tried a number of AI sites and found that they do not read all the pdf to give an answer. I have done some testing and my theory seems to be true.
After a news email I received about GPT4 is that:
It can handle 25,000+ words from the previous 3,000 in GPT-3.5.
Does this mean what I think it is regarding how much GPT actually reads for and answer.
Thanks for any replies
The 32K token limit (~24K words) feature for
gpt-4 has not been released yet, to my knowledge,
I have used a few of the ‘pdf reader’ or ‘chat with your pdf’ plugins, and I find them incredibly frustrating.
The one that best access a file from the web is askyourpdf, but even then, it doesn’t answer questions about the entire document. I uploaded a 640k pdf story to the askyourpdf plugin, got the doc ID, used the plugin with the doc ID.
Then using the plugin, it will tell me about some of the story, but not all of it… for example, there is a character named Jim who is in 80% of the story, but the plugin says there is no such character. It also refers to ‘excerpts’ so it makes me believe that the whole document is not loaded.
This seems like one of the MOST BASIC use cases for an AI, feed it a document and then ask it about the document and have it use the doc as an information source, this should be core functionality from OpenAI and not foisted off onto janky plugins.