When it comes time to index these vectors for better performance, you have two options. You have HNSW and IVF flat, and they're slightly different, and we'll talk about that in a second. Indexing vectors breaks one of the cardinal rules of indexing, which is the index should not change the results of the query. That is not true when it comes to vectors. When you are running a query against a vector, embedding without an index, it's performing an exact nearest neighbor search. When you're running that query with an index, it's performing an approximate nearest neighbor search, and you have control, you can tune some of the parameters to control that trade off.

It's not necessarily a bad thing, it is just a thing you need to know. When you add an index, it might come back with a different set of results, which would be very surprising to me if I did not know that, because indexing shouldn't change the results. In this case, it can. That's a trade off that is mostly, we're mostly willing to make for speed.

Let's take a look at these two types of indexes. The first one we're going to look at is probably the best general purpose. It does have a few drawbacks, but we're going to do create index on articles. We'll create index, let's say articles, embedding IVF flat, inverted file flat, which doesn't make a lot of sense to me, but that is the name of it. We're going to say using IVF flat, and then we'll say embedding. Then you can pass a flag here. You can also say with lists equal something. There are a few things. There are a few options that you have here. The first is lists, and this is, it breaks it apart into a certain number of lists, and there's a formula to calculate the general, I guess, best practice for lists. That's up to a million rows. It's the rows divided by a thousand, and then after a million rows, it's just the square root of the rows. That can kind of help you with the lists. Now when you create an IVF flat index, it's best if there's already data in the table. That's a little bit of a drawback.

Then what we're going to do now is we're going to pass through the operator that we intend to use with this index. Right here after the column name, we're going to put vector_l2_ops, and that says we plan to use the l2 operator here. Of course, you could do IP for inner product, you could do co-sign for co-sign, and this is going to tell Postgres or pgvector rather. This is the operator that we're going to use. Please use this index when we use this operator.

The other type of index, the other type of index is called the HNSW. We'll just add HNSW as the end of that, and we'll call this HNSW, and that's what creating an HNSW index would look like. You also have the same option here. You can change it to l2 or inner product, whatever. HNSW is a little bit of a higher quality index you can get some better performance and better accuracy, better recall with an HNSW, but it is a little bit more expensive to maintain. It requires more resources, both in memory and compute. A good general purpose is the IVF flat, but if you're willing to do it, the HNSW can be a little bit higher performance. There are lots of parameters that you can tune on these things, on these indexes, resource allocation recall and accuracy, performance, all of that kind of stuff. Lots of parameters both at the index build time and at the query time.

Now, this is all still very, very much in flux. PG Vector is not even at a 1.0 release at the time of recording. It's at a 0.8, I think. Everybody's using it and it's not even 1.0. It's all changing very, very quickly. As much as I would love to tell you this is exactly what you should do in this case, that's kind of hard. That's kind of hard for me to tell you exactly what you should do here. I'm hoping, hoping that I've given you some right down the middle use cases, some generalized principles, but because of how fast, not only this vector embedding, but also PG Vector itself because of how fast all of that is changing, I'm loathed to give you very strict rules here.

Hopefully I've given you some general principles here. IVF Flat is a great general use case index if you are willing to commit the resources. HNSW can be higher performance and better accuracy at the cost of being a little bit more expensive. If you're just going right down the middle, hopefully now you're able to go out and build something today.