🤖 RAG on AWS: Building an AI-powered Knowledge Base, with Amazon Bedrock and S3 Vectors

🏃‍♂️ TL;DR

• AWS released Amazon S3 Vectors as native vector storage inside S3.
• Store, index, and query billions of vectors with sub-second latency.
• Up to 90% cheaper than traditional vector DB setups.
• Integrated with Bedrock Knowledge Bases, SageMaker Studio, and OpenSearch out of the box.
• Still in preview! No CloudFormation/CDK support yet, so it's not ready for core prod systems but a perfect playground for builders who want to experiment with AI-ready storage.

🚀 Rethinking how we store and query vectors

If you read the first article in this series, I've explored how to build a RAG pipeline with Amazon Bedrock Knowledge Bases using Pinecone. The reasoning was simple: Pinecone is a vector database designed for AI, natively integrated with Bedrock, and way more cost-effective than running Amazon OpenSearch just for embeddings.

But today, I’d like to talk about something new that could completely change how we think about vector storage: Amazon S3 Vectors.

If you’ve been building AI agents, semantic search, or anything that relies on embeddings, you already know the story: vectors are everywhere. But storing, indexing, and querying them at scale?

That’s usually been a pain: costly, complex, and often involving extra infra you don’t really want to handle.

That’s where Amazon S3 Vectors comes in.

🔍 What is actually S3 Vectors?

S3 Vectors is the first cloud object store with native vector support. Basically, Amazon S3 now has built-in APIs to store, access, and query vectors directly.

Why this is a big deal for builders?

1. 90% cost savings compared to traditional vector databases (uploading, storing, querying).
2. Sub-second query performance, even at massive scale.
3. S3 durability and elasticity
4. AI-native: purpose built for AI agents, semantic search, and RAG.

💡 Build faster with AI-ready storage

What I really love about this new S3 option, is the out of the box integration for Amazon Bedrock Knowledge Bases (among others) which makes Retrieval Augmented Generation (RAG) way simpler and cheaper.

A picture is worth a thousand words (credits to awesome article "Introducting Amazon S3 Vectors)

What about Amazon OpenSearch service or solution as Pinecone?
You could tier your vector data:

• keep the “long-term memory” cheap in S3
• while “short-term memory” hot in Pinecone/OpenSearch for fast inference.

This combo means you don’t have to choose between cost-efficiency and performance. You can choose the best in class for your use case.

🚀 Create a S3 vectors-powered RAG with Amazon Knowledge Bases

First of all, go to Amazon Knowledge Bases console, click on create button then select vector option

As second step, give our Knowledge Base a name as we are familiar.

You should now select a source, let's go with standard S3 object storage. We'll store some csv files here as document source for our RAG.

You should now create a S3 vector store

or select a previously created S3 vector store

Finally, just review your selections and create an Amazon Bedrock Knowledge Base

Here is the section of S3 Vector Store

🧪 Test it out

You can simply test your RAG powered by your newly created vector store. Let's start uploading some file to your source standard S3 bucket.

Then sync your Amazon Bedrock Knowledge Base and try some relevant question for your data: as I've added big mac cost around the world and Tokyo Olympics medal results, I'm asking some simple question about it.

You can easily review details of retrieved data in the test panel

⚠️ Things to keep in mind

As exciting as S3 Vectors is, it’s still in preview. That means:

• Not production-ready (yet): it’s awesome for experiments, prototyping, and side projects, but I wouldn’t bet the core of a production system on it right now. Expect some rough edges and possible changes in APIs or behavior before GA.
• No CloudFormation/CDK integration (yet): this is a big one. Right now, you can’t just spin up S3 Vectors resources via Infrastructure as Code (IaC). For builders who rely on repeatable, automated deployments, that’s a blocker for serious production adoption. Once CloudFormation and CDK support land, that’s when I think we’ll see this become a mainstream building block in real world AI projects.

📌 Final thoughts

If you’re a builder, now’s the perfect time to experiment and get familiar with S3 Vectors. But if you’re running a mission critical app, you should treat it as a preview: learn it, play with it, and be ready to adopt when the full production tooling support arrives.

For me, this feels like one of those “AWS building block” that changes the game and I’m already thinking about how to re-architect some of my RAGs to cut costs and simplify cloud infrastructure.

🌐 Resources

You can find some useful resources about S3 vector here and here.

Moreover, you can find here a useful cli to interact directly with your S3 vector: don’t miss the API to query with metadata as it’s super handy!

🙋 Who am I

I'm D. De Sio and I work as a Head of Software Engineering in Eleva.
I'm currently (Apr 2025) an AWS Certified Solution Architect Professional and AWS Certified DevOps Engineer Professional, but also a User Group Leader (in Pavia), an AWS Community Builder and, last but not least, a #serverless enthusiast.

My work in this field is to advocate about serverless and help as more dev teams to adopt it, as well as customers break their monolith into API and micro-services using it.