So last Friday I sent you and email about GPT-5 and how you can stay ahead of the game in this rapidly evolving field.

One thing I didn't have the time to expand on was

Knowledge Bases

These are basically additional memory for ChatGPT, and allow us to seemingly bypass context window limitations.

In essence, what these allow us to do is to upload lots of files to a database, and then conntect an LLM like GPT-4 to it.

These work similar to you searching your files for a keyword. Let's say you remember you had a research paper quoting the "effects of a lack of sleep on the brain and mood regulation" but can't remember where you put that file.

So you open up Finder or Windows File Browser (or press Cmd+Space on a Mac if your files are indexed with Spotlight) and search for these words.

And depending on how well you search and what keywords you pick it will take you more or less time to find the document.

Once you have the document, you need to find the relevant part that talks about this.

If you need a direct quote, you can stop here.

If you want to add this text as extra context to ChatGPT, you can select the relevant part and copy-paste into ChatGPT and make it do something with it.

This is what these knowledge bases can do, but they can do it at scale.

It's also called RAG (Retrieval Augmented Generation), which means you are Generating text (with an AI) but you are using a Retrieval system to Augment the response, and give it extra context.

I don't want to get into the nitty-gritty of how these work, other people have done that before me. (This is Pinecone's official article and it's super technical.)

The point of this email is to give you a perspective on what's possible with this.

You can upload all of your internal documents to such a software (our top 3 are below), and basically "train" the AI model to know the information in them.

It's not the same as adding more training data to ChatGPT and actually training it, but it feels like it, because at the end of the day, what happens is that ChatGPT can now give you answers from your own documents!

Without a knowledge base, you have to manually find the information from your documents and give them to ChatGPT.

With a knowledge base, you can just upload all stuff in an unstructured way and it will find the relevant information for you.

Solution 1: Keymate

With Keymate, you can use your knowledge base from the Keymate Webapp or within a Custom GPT so it's even more convenient.

This makes it easy to upload the files you want ChatGPT to use, and you can always just call the Keymate GPT into a chat (with the @ symbol within ChatGPT), ask it a question that needs your knowledge base, and then take it from there.

Keymate also uses Google search instead of the Bing search that's built-in to ChatGPT, and Google seems to provide better results.

However, if you are looking to manage multiple knowledge bases (maybe per different topic or per client) there is another solution:

Solution 2: Superpowered.ai

With Superpowered.ai, you can set up multiple knowledge bases, define which one you want the model to use and even define how it retrieves the information.

Superpowered AI also has an advanced Indexing system, which gives it the ability to remember citations, so it can tell which document it got the information from.

Solution 3: OpenAI embeddings

These are integrated into the Custom GPTs and Assistant APIs, so whenever you upload a document to a Custom GPT, it gets embedded and the GPT can read it through retrieval (if prompted properly).

In Custom GPTs, these are limited to 10 files only, but no limit to how big a file can be, so easy to bypass by merging PDFs into one giant PDF.

So my suggestion is to learn and integrate these into your processes, but don't bother developing new RAG solutions. The ones we have are already good enough, fast enough and cheap enough.

Sure they have some limitations (like being text-only, and not being able to handle images, charts, tables unless what you see is described in text format - or maybe by GPT-4Vision?), but overall, they help overcome a big limitation of LLMs: Not having information on certain topics/languages, after the training date or private documents.

Once you master this, clients will be standing in line to get help integrating a custom knowledge base solution into their company processes.

For example, to make their chatbot answer based on their internal docs, to make an AI Agent that answers support tickets, to create marketing content using their own IP, and so much more.

And the cool thing about it is that RAG is here to stay. It seems like a future-proof skill for a while, as even if GPT-5 supercedes GPT-4 by the same jump that's between GPT-3.5 vs. GPT-4, GPT-5 won't have access to your own docs.

Even if Claude-4 is better than GPT-5 it still won't have access to your own docs. RAG is a totally different technology.

I hope this email helped you better understand what this technology is about and how it fits within this whole AI landscape.

See you on Wednesday,

Best,

Dave Talas, COO, co-founder of Promptmaster

P.S: Click here to become a Prompt Master and make ChatGPT work for you — at a pre-order price!

P.P.S: If you have burning questions you'd like us to answer, book a consulting call here.

‍