"And that brings us to Retrieval Augmented Generation. Let's imagine it as your model's library card. So let's go back to tokens. Remember what we learned about tokens? They're pieces of input that you feed into a model. Models have token limits. Sometimes you know the model and you've got lots of information you want to reference, but your model just isn't going to be able to fit all of it into its context window. It also doesn't have any of that information. Maybe these are a big stack of recent newsletters. It was not trained on that information. This is all news about things that are going down around the world. How are you going to be able to reason about that news with your model? You can't just shove it into that context window. It won't fit through the door. When that happens, the first thing you do is you embed the new data into a vector database. You put all of those reference materials into your library where the information is stored near similar information. So you've got all your newsletters, all of your new friends in there, and you want to ask a question. I would like to go on a vacation with my friend. What are some good places we could go to? So that query gets logged into the vector database. It gets set up and it's near the other information. So next, the LLM can check out the related books by getting just the related embeddings needed to process the query in its context window. Hurrah. That was a tough one. I really wanted to explain HRAG to the layman because I felt that it was being proposed as a solution to far too many things."