I asked GPT-4 to compare Regular ChatGPT and Custom GPTs in terms of their retrieval approach, and this is the comparison table it provided:
Is this table correct?
Hi @telem.yuval and welcome to the community!
My understanding is that Custom GPT is exactly the same as “regular” ChatGPT. The only difference is that with Custom GPT you can (to a large extent) confine it to some specific knowledge, e.g. by uploading certain data, documents, text files, images, etc, and by specifying certain roles/instructions/guidelines. So anything “regular” ChatGPT can do, so can custom GPT in the same way. To my knowledge there is no RAG or vector search in neither of them (this is something that is instead used in Assistants API). The only addition that custom GPT provides in terms of capabilities is that you can define “actions”, e.g. by providing specifications on how to talk to certain APIs and external services.
So it’s best to think of custom GPT as ChatGPT but confined to a specific role/task with specific data access.
I’m just starting on the custom GPT road through the UI (no APIs yet). I have a lot of technical documentation that I use internally and I’m making a techadivsor GPT for my field engineers. There is a limit to uploading 20 files. If I understand correctly, that’s what you get, 20 manuals at 512mb per. A specific user session, you can upload more, but they don’t get stored for long and are only available in that session? If there is any strategy to having smart internal GPTs that you can share, I would greatly appreciate it.
If the gross mb of the files you need to provide the GPT exceed the limits, I suggest you preprocess and extract the relevant data from them and create file types that will inherently be text based, and therefore smaller. I typically extract data from .csv, .pdfs, and .xlsx files and generate .json or .md files for upload to the GPT. It’s relatively easy to do - I use perplexity to generate the python code (in my experience GPT4o is quite poor at python coding) to extract and format the data. You’ll need to iterate a bit to get it just right, but it’s worth it. The other thing I do is combine .json or .md files so that I can use the maximum file size when necessary.
I appreciate that. It’s the total number of files. The size isn’t the limiting factor but as you point out plain old ASCII is smaller and easier to parse.
I haven’t programmed in decades but I’ll give it a go. Thank you!
Hi @platypus , Thanks for your reply! I just came across this article that mentions RAG with GPTs.
What’s your take on this?
