#How do I find relevant insights between 2 large documents?
23 messages · Page 1 of 1 (latest)
Just use GPT4 API and give relevant promts
@coral vale I dont have access to 4
You can join the waitlist, either I was testing a way to use GPT3 as 4 by giving multiple promts, but it uses alot of tokens, if that's fine, you can consider this
I just applied for waitlist, but I'm also going to operate as if I won't get access for 2 more years, so I am building a solution now and looking for advice regarding my current capabilities with 3.5
Yeah you can actually use 3.5 instead of 4 and bypass the text limit, just we have to use multiple promts by summarising the previous one and sending it to the new one, it will cost you alot though but it's the best option to bypass text size limit by using 3.5
I can help you with it
@coral vale what about embeddings as a means to a solution?
lets assume also for simplicity sake that the sum tokens of document A+B are under the token limit.
Yeah embedding will work fine, but some details can be lost as we have to compare the documents so it might not be the best practice
I see. Then assuming sum tokens of document A+B are under the token limit what technique would. you advise? I tried a quick demo using the chat.openai.com and it was confused by the prompt.
First describe what gpt is, like a highly trained and professional legal advisor or something, and then give a situation which creates a sense of importance, then provide it the documents with some brackets so it is differentiated by rest of the test, and then finally tell what he have to do like 5 questions about the legal document A given the situation described in legal document B
it may have gotten confused by the lack of "" and its possible the documents had quotes in them in the first place
my issue is these documents are thousands of tokens
I would rather put document A into "memory" and document B into "memory" and then my prompt does not need to be 4k characters long, but I suspect this may not be possible with 3.5?
That's why I recommended brackets to differentiate
It is possible, the way I told you earlier, it just requires multiple promts that work with each other
is there an existing example of this multiple prompting?
Both will work but just make sure same sequence is not in the document
I have find and used this for my own personal project, which is not publically available
Ok, could you please summarize the technique maybe in a few pseudo code lines?