I'm creating my own GPT based on a few books. I learned someone has done this by uploading their books into ChatGPT. They have written around 25 books or so. Some are best sellers so when people go the this GPT and ask questions based on the topic, the AI only pulls answers based on the books. What's the most effective way to do this? In chunks, say one chapter at a time? The whole book at once? Or am I best to get ChatGPT to break the book down then use that info to feed my GPT? Thanks in advance for your help.
#Custom GPT.
wicked talon
high agate
You can add knowledge files to a Custom GPT when setting it up. I believe there is a limit to the number of files, file size or a combination, so you may not be able to scale it to the size you want but this will also likely dictate whether you can in fact upload the chapters as their own separate files or in more of a batched format.
It's probably a good thing to be aware of how the knowledge is ingested and used. With this set up the Custom GPT will not necessarily understand all of the content from the initial chat message/query, but will have to open and read the files as and when it is able to match up the users query with the summarised header it has tagged those files with at setup.
You may have to wrestle with the Instructions you apply to the Custom GPT to ensure it provides answers from the chapters/books as LLMs generally have a very high preference to answer based on its own intuition rather than refer to dedicated documentation. I think this is probably unavoidable at some level as it would have to trust its own intuition to be able to form coherent sentences in the first place, but I digress. You'll need to research or experiment with the instructions.
If you want to learn more about its memory search for the topics of context length, RAG, fine tuning. Custom GPTs can also make use of a web API to extend their capabilities further, but this'll probably be beyond the scope here.
normal marsh
Use this at the end of the main instruction block, I hope it helps:
Important Review the knowledge files in each user interaction.