Vector database

content.

So then that whole map lives in what's called a vector database.

And it feels like a really complicated term, but basically it's just a database.

And so when a customer types in a question, that question kind of gets oriented on a map and then the system grabs the chunks sitting close to it.

So the ones that are the closest in meaning and then hands that to the AI.

So basically when things are embedded, they get grouped together with similar content and They live in a vector database.

And then that's how the AI is able to go and find what it's looking for.

Yeah, and the whole reason it can find your returns policy or a specific passage in a document is because when someone's asking about, you know, when someone asks that question, it's not that it's matching against the words.

I think that's the pattern that's interesting.

Think about vectors as almost the precursor to this.

What vectors was really making you do was, it was making it easier for you to search, retrieve, and reason on the data you already had.

It was turning the data that you had into embeddings or numerical translation so that when you're doing pattern matching against the inbound query, you now know which documents or files or records to go retrieve instead of trying to peruse every single one at the same time.

Vectors were a very primitive version of an index that you can think about, but you can get so much more sophisticated than just having vectors to match the document because you can actually derive signal from the content that's coming in there.

To your point though, these are techniques that have existed in search for a very long time.

When you are doing web search, this is going back to even my days at Microsoft when I was working with Bing, you would crawl tens of millions of web documents and extract things like This page is about Taylor Swift and these images exist.

They're creating more memory, they're creating more context, that's going to be a layer that always continues to evolve.