Amazon S3 Vectors + Bedrock: RAG at 90% Lower Cost

Category: Cloud Computing

Tags: AWScost-savingdatabasesS3vectors

Entities: Amazon BedrockAWSOpenSearchRickyS3 bucketsS3 vectorsTitan embeddings

Transcript

00:00

Hi, this is Ricky from data mastery and in today's video we are going to talk about S3 vectors. So S3 vectors is a fairly new feature or a service which was recently announced by AWS and it is

00:15

not GA right now. It is still in preview.

So uh be very cautious if you want to use this in prod. But in today's video, we are going to understand what is S3 vectors and how we can use that in our rag workflows.

00:33

So as you probably already figured out S3 vectors is based on the S3 service and instead of using very costly vector search or vector store databases like open search you can simply use S3

00:48

buckets uh as your vector search or vector store databases and since it's backed by S3 you can definitely see cost reduction around 80 to 90% as compared to open search or any other vector store database. So if

01:06

you're building rag workflow and if you're on AWS, this is fairly kind of uh easy integration that you can have in your rag workflows. So let's go ahead and see how we can use S3 buckets as our vector stores.

01:22

So I'm logged into my AWS account and this is Amazon Bedrock service. I'm going to go ahead and create a new knowledge base.

But before I do so, I will take a moment to go to S3 and I will show you a bucket which is already

01:39

there that contains some of the documents that I'm going to ingest into S3 vector. So I have a bucket called data mastery can draw demo and I have few PDFs here.

Code of conduct policy, employee

01:55

benefits and a vocation policy. So I'll go ahead and create a new knowledge base.

And when you create a new knowledge base, you have to select knowledge base with vector store. And here you can give any name to your

02:11

knowledge base. I will just keep is this random name that it generated automatically.

And then you have to select the IM role. If you have existing IM role, you can select that or you can create a new uh service role.

And then

02:26

you have to select your data source. Now this is the data source where your data is like in my case I have this bucket.

This is my data source. This is not the vector store.

So choose S3 as your data source or any

02:42

other data source where your data is. And click on next.

So then we have to give a name to our data source. You can give any meaningful name or just keep the randomly generated name.

And here we have to specify uh the

03:00

data source bucket. So I just going to select my data mastery Kendra demo bucket and I will keep other options as is.

Now I have to select the embeddings

03:16

model. Embeddings model is required to create the embeddings and store in your vector store.

And here I'm going to pick the Titan embeddings from Amazon. For the vector store, now you have two options here.

You can select a quick

03:34

create uh for this PC. I'm just going to use this first option.

And if you look at the options here for the quick create vector stores, you have all the options which are Amazon or AWS services. So you have open search and here you see Amazon

03:50

S3 vectors as a new option. You have Aurora Postgress, you have Neptune Analytics.

And if you want to pick an existing vector store, for example, if you have already created a vector store bucket, then you can pick that. And in the

04:06

drop-down list, you can see that you will get an option of S3 vectors. And here you have to specify the bucket ARN and the index.

So I'm just going to quickly create a new vector store. And I'm going to

04:24

select S3 vectors. Click on next.

And you can review all of your options here. And then click on create knowledge base.

Okay. Now the knowledge base has been created.

And this is my data source. If

04:40

I click on this data source and you can see that we have not indexed or synced it with actual data. So what I'm going to do now is I'm going to click on sync.

So it is going to sync the data from my data source bucket. And you can see that

04:55

the sync was very fast and it has already completed. I had three files and all the three files have been added.

And if I just refresh it once

05:12

and here you see these three documents, these three PDFs are indexed. So basically what happened the data which was there in my data source bucket the three documents those are synced and indexed and the embeddings are generated

05:28

and stored in the vector store bucket. Now if we go to S3 you see this new option here called vector buckets.

Now this is not a generalpurpose bucket it's a vector

05:44

bucket. So you can click on vector buckets and you see that a new bucket has been created for you and if you click on the bucket and this is the default index which has been created.

Now this index stores the embeddings about those three PDFs that we have

06:00

ingested. Now let's go back to the knowledge base and if I click on the knowledge base you can see all the details of the knowledge base the data source name and if you scroll down you can see the embeddings model that we have selected and also the Amazon S3

06:17

vector as our vector store and you can see the ARN of the vector store bucket which is what we have seen here already. Now once you have your knowledge base created and your data source has been synced, you can go ahead and test your

06:34

uh knowledge base right in the console. So there's this link here called test knowledge base.

You can simply click this and here you have to select the model that you want to use for query and response. You can go ahead and click on select model and let's say I'm going to

06:51

select entropic and I'm going to select claw 3.5. You need to have the access to these models first.

If you don't have the access to these models, you can go ahead and click on this model access link and then you can request for access for the

07:07

model that you want to use. So I'm going to keep all other options as is.

I'm not going to change anything. And here I'm just going to ask a question based on the documents that I ingested.

So let's say I'm going to ask can you summarize

07:24

the uh let's say vacation policy. So here you can see that the vacation policy has been summarized and you can see that the employees are entitled to 20 working days of paid vacation per

07:40

year. And if you click on the details you will also see where this data is coming from.

So you can see this is coming from the vacation policy PDF. So this is working pretty good.

And you can

07:57

go ahead and ask another question. Let's say summarize the health benefits.

08:14

And here it is summarizing the health benefits for you. And if you click on details again and you will see that this information is coming from employeebenefits.pdf.

So this is working pretty good. Now if you look at this workflow it is very

08:30

similar to your normal rag workflow like there is nothing fancy about this. The only change that we have made in this case is to use s3 vector store instead of open search vector store.

And this is going to help you save a lot of money

08:46

almost about 90% as compared to open search. Now one more thing I would like to show you is how you can let's say delete the uh the vector store buckets.

So, if I go ahead and

09:03

click on my knowledge base and I'm going to delete it. So, it's going to delete the knowledge base.

It's going to take a few minutes,

09:19

but it will not delete the S3 vector store bucket. And if you go to S3, you don't see any option to delete the vector bucket.

So how you going to do that? If you have to clean up the vector

09:34

buckets. Well, for now you you can do that using the API.

So if I let's say open my cloud shell. Now this is not a normal S3 bucket.

This is not general purpose. So AWS has released a new API for S3 vector which

09:53

is called S3 vectors. So let's see how that looks like.

So I have AWS S3 and if I just hit on tab you can see that we have S3 we have S3 API we have S3 control and S3 vectors

10:11

is the new one for the vector buckets. So click on S3 vectors and let's say I'm going to say list vector buckets.

And here you see this is my vector

10:29

bucket that you can see in the console as well. And if I have to list indexes and the vector

10:45

bucket name is going to be this. And you can see this is the index ARN.

And if you click on this basically it is

11:02

referring this index arn right here. Now in order to delete or clean up this vector bucket first we have to delete anything which is there in the bucket.

So right now we have the index which is

11:18

in the bucket. So if I let's say go ahead and say delete index.

So I have to provide index name which is

11:35

this and the vector bucket name which is this.

11:54

And here you go. If you refresh you can see that there's no index anymore.

And finally, you can clean up the vector bucket as well. Say delete vector bucket.

12:11

Vector bucket name. And here you go.

So you go back to your vector buckets.

12:27

And now you can see that there is no vector bucket anymore. And go to the knowledge bases.

So it is still getting deleted. It's going to take a few minutes before uh it finishes.

But keep in mind one thing that you will not be able to delete the

12:42

vector buckets from the console as of today as of the time of recording of this video. Since this is in preview, you can expect more functionalities around this both in API as well as console.

12:58

But however, the takeaway uh from this video is that you can use S3 vector store instead of the costly options like open search and other vector store databases and you can immediately cut the cost by 85 to 90% for your rag

13:16

workflows. So I hope you find this video helpful and whenever you get a chance you can just play with this.

If you have any question whatsoever, you can ask that in the comment section. And thanks again for watching the video.

Do subscribe to our channel if you haven't

13:32

done that. And I'm going to meet you next time with a similarformational video.

Thanks for watching. Cheers.