Noelle Russell explains why AI models make up plausible-sounding details, then shows how Retrieval-Augmented Generation fixes it by connecting the model to real data sources like SharePoint. Instead of guessing, the model retrieves actual documents and gives you facts instead of hallucinations.
Chapters
as model errors. When the model makes a mistake in a generative model, that mistake often means it's going to, uh, creatively fill in the gaps that it doesn't fully understand. I typically use an example of like someone coming up to me and asking me to, to recall what I did in 9th grade in my English class, like the paper I wrote in 9th grade, right? And what I would likely do is kind of as a human look back into 9th grade and be like, okay, I kind of sort of remember 9th grade. I kind of remember my teacher. I'm going to be very good at certain details that I remember, but the rest is going to get fuzzy. And I'm just going to kind of make it up typically as a human. I'll just fill in the gaps and create a pretty plausibleโ like, I think I probably wrote a paper on, you know, I remember writing a paper in 9th grade on technical writing, I think it was. Mm-hmm. So I can speak with confidence, which some of us do, speak with confidence that I wrote this paper even though As I get more detailed in the answer, I have to kind of make it up because I don't exactly remember. And that's kind of how an LLM works, is that it's going to be able to start off pretty good. But as you get more and more specific, it's going to fade in its awareness and understanding, which is where retrieval augmented generation comes in. Because now, though, if I wanted to be really good at that question, I'm going to give it all my papers from 9th grade, right? I'm going to source those papers in a SharePoint directory and say, when you get asked questions about this, Go to this SharePoint directory, right? If I had an old-fashioned file folder, which I do, of all the papers I had ever written, I would go to that folder, pull out the manila folder, and be like, oh, here it is. But that system of recollection has to come from somewhere. So that is one problem that retrieval augmentedโ augmentation solves, is to source for a model, ground a model in a data source. so it knows where to get the answer instead of using inferencing on its training data, which is where that fuzziness comes in.
