Understanding Basic Vector Search With KNN | Vector Databases for Beginners | Part 12

Key Takeaways

So the main algorithm I'll start with is KN&N or K nearest neighbors. And this is when you're calculating the distance between a query vector and a document vector for every query and document pair. So, we covered this in the last webinar, but when you turn a query or a document or anything into a vector, you're basically encoding its meaning into this multi-dimensional space. From that sort of string of numbers that represents its meaning, you can do calculations, mathematical calculations that can then represent how similar or not similar two vectors are. And so we can use this in vector search to basically find the most similar documents to a query. So what sort of happens behind the scenes in this is that you will store all your document embeddings. You'll embed them with an embedding model and store all of them in a database and then you'll convert your query when you get it to a vector embedding. And then you can calculate a similarity score for all your documents to the query by measuring the distance between every document in the query. And then basically you'll just return the documents with the highest score. So we can represent vectors in sort of a three-dimensional space. They're multi- many more dimensions than three dimensions usually. But you can see here like if our query was kitten, the most similar vectors would be like cat and dog and wolf. The vectors that are closer in distance, more different vectors like fruits would be farther apart. So closer distance represents higher similarity. And in this way we can do vector search by measuring distance between vectors. There are several different distance metrics that you can use in vector search and I'm not going to go into detail in a lot of them. We have a nice blog post on this if you're interested in details. But what I want to get from this is it's math, right? You're taking a vector and you're doing math things to it. And in this way, you're representing similarity with a score with a math score. Um, and that's basically it. So, keyword search is going to return results that have the exact keyword match, right? So if you're searching for cola, it's going to return all the Coca-Cola products. Vector search, on the other hand, is going to return results with related meaning. So if you search from cola, maybe it'll also return Pepsi or Fanta or things like this. Similarity through meaning, not from exact keywords.