At the bottom, in their coming soon section: "Customization: Securely extend ChatGPT’s knowledge with your company data by connecting the applications you already use"
- Codebases
- Documents (by way of connection to your Box/SharePoint/GSuite account)
- Knowledgebases (I'm thinking of something like a Notion here)
I'm really looking forward to seeing what they come up with here, as I think this is a truly killer use case that will push LLMs into mainstream enterprise usage. My company uses Notion and has an enormous amount of information on there. If I could ask it things like "Which customer is integrated with tool X" (we keep a record of this on the customer page in Notion) and get a correct response, that would be immensely helpful to me. Similar with connecting a support person to a knowledgebase of answers that becomes incredibly easy to search.
Fine-tuning isn't great at learning knowledge. It's good at adopting tone or format. For example, a chirpy helper bot, or a bot that outputs specifically formatted JSON.
I also doubt they're going to have a great system for fine-tuning. Successful fine-tuning requires some thought into what the data looks like (bare docs won't work), at which point you have technical people working on the project anyway.
Their future connection system will probably be in the format of API prompts to request data from an enterprise system using their existing function fine-tuning feature. They tried this already with plugins, and they didn't work very well. Maybe they'll come up with a better system. Generally this works better if you write your own simple API for it to interface with which does a lot of the heavy lifting to interface with the actual enterprise systems, so the AI doesn't output garbled API requests so much.
I do think an open line of research is some way for users to just add arbitrary docs in an easy way to the LLM.
I certainly don't expect a nice drag-and-drop interface to put my Office files and then ask questions about it coming in 2023. Maybe 2024?
Unfortunately even if we do get this, I expect there will be significant ecosystem lock-in. Like, I imagine Microsoft is aiming for something like this, but you'd need to use all their stuff.
Rather than wanting to import N documents per month, I would want to import M documents all at once, then use that set of documents until at some future time I want to import another batch of K documents (probably a lot smaller than M) or just one document once in a while.
By limiting it to a fixed amount of documents per month, it eliminates all the applications where you need to import a complete corpus before the service is useful.
Anyone knows how this new capability works in terms of where the model inference be done? Would it still be at the OpenAI side or is this going to be at the customer side?
After using RAG with pgvector for the last few months with temperature 0, it's been pretty great with very little hallucination.
The small context window is the limiting factor.
In principle, I don't see the difference between a bunch of fine-tuned prompts along the lines of "here is another context section: <~4k-n tokens of the corpus>", which is the same as what it looks like in a RAG prompt anyway.
Maybe the distinction of whether it is for "tone" or "context" is based on the role of the given prompts and not restricted by the fine-tuning process itself?
In theory, fine-tuning it on ~100k tokens like that would allow for better inference, even with the RAG prompt that includes a few sections from the same corpus. It would prevent issues where the vector search results are too thin despite their high similarity. E.g. picking out one or two sections of a book which is actually really long.
For example, I've seen some folks use arbitrary chunking of tokens in batches of 1k or so as an easy config for implementation, but that totally breaks the semantic meaning of longer paragraphs, and those paragraphs might not come back grouped together from the vector search. My approach there has been manual curation of sections allowing variations from 50 to 3k tokens to get the chunks to be more natural. It has worked well but I could still see having the whole corpus fine-tuned as extra insurance against losing context.
Other considerations: (A) would you fine-tune daily? weekly? as data changes? (B) Cost and availability of GPUs (there's a current shortage)
My experience is that RAG is the way to go, at least right now.
But you have to make sure your retrieval engine work optimally: getting the very most relevant pieces of text from your data: (1) using a good chunking strategy that's better than arbitrary 1K or 2K chars (2) using a good embedding model (3) Using hybrid search, and a few other things like that.
Certainly the availability of longer sequence models is a big help
Sharing this relevant discussion from LinkedIn: https://www.linkedin.com/feed/update/urn:li:activity:7101638...