In this two minute video from my How AI Ate My Website talk, I outline how to automatically answer people's design questions using the content from Web site using embeddings. I also explain why that approach differs from how broader Large Language Model (LLM) generate answers. It's a quick look at how to make use of AI models to rethink how people can interact with Web sites.
When we have all these cleaned up bits of content, how do we get the right ones to assemble a useful answer to someone's question? Well, all those chunks of content get mapped to a multi-dimensional vector space that puts related bits of information together. So things that are mobile-touch-ish end up in one area, and things that are e-commerce-ish end up closer to another area.
This is a pretty big simplification, but it's a useful way of thinking about what's happening. To get into more details... enter the obligatory system diagram.
The docs that we have, videos, audios, webpages, get cleaned up and mapped to parts of that embedding index. When someone asks a question, we retrieve the most relevant parts, rank them, sometimes a few times, put it together for an AI language model to summarize in the shape of an answer.
And sometimes we even get multiple answers and rank the best one before showing it to anybody. Feedback is also a really important part of this, and why kind of starting with something that roughly works and iterating is more important than doing it exactly right the first time.
So what's the impact of doing all this versus just using something like ChatGPT to ask questions?
Well for starters, you get very different kinds of answers, much more focused and reflecting a particular point of view versus general world knowledge. As you can see in the difference between a ChatGPT answer on the left to, why do designs look the same, versus the answer you get from Ask Luke.
On the Ask Luke side, you also get citations, which allow us to do a bunch of additional things, like object-specific experiences. On Ask Luke, you ask a question, get an answer, with citations to videos, audio files, webpages, PDFs, etc. Each one has a unified, but document-type specific interface.