Zero-Infrastructure Vector Search with Amazon S3 Vectors

The world of generative AI is evolving at a rapid pace and one of the most powerful and practical applications is Retrieval-Augmented Generation (RAG). RAG enhances Large Language Models (LLMs) by giving them access to external, up-to-date knowledge bases. This allows them to generate more accurate and context-aware responses.

Traditionally, building a RAG system required setting up and managing a separate vector database that adds complexity, cost, and a new layer of infrastructure to maintain however with the introduction of Amazon S3 Vector Buckets a new paradigm has emerged: zero-infrastructure vector search.

What is Zero-Infrastructure Vector Search?

Amazon S3 is a highly scalable, durable, and cost-effective object storage service.With its new native vector support (preview), S3 is no longer just a passive data lake. It now functions as a serverless vector search engine. This means you can store, index, and query vector embeddings directly in your S3 buckets, eliminating the need for a separate and dedicated vector database.

The key benefits of this approach are:

Cost-Effective: It can reduce the total cost of uploading, storing, and querying vectors by up to 90% compared to other solutions as you only pay for what you use.
Simplified Operations: There’s no server provisioning, patching, or scaling to manage. S3 handles all the underlying infrastructure.
Scalability: S3 is designed to handle massive volumes of data from millions to billions of vectors effortlessly.

This makes S3 Vectors an ideal solution for RAG applications that don’t require millisecond-level latency and can tolerate sub-second query performance in exchange for significant cost savings and operational simplicity.

Create a Vector Bucket and Understand Vector Indexes

Head over to S3 in the AWS Console: Once you are logged in, navigate to the S3 service.

Find the Vector Buckets tab: In the left-hand navigation pane, find and click on Vector Buckets.

Amazon S3 Vector Bucket Create Vector Bucket

Create a Vector Bucket: In the dashboard, click on Create vector bucket. Give your bucket a unique name and click Create. You can leave the other settings as their defaults for this example.

Vector Indexes

A vector index is a resource within a vector bucket that stores and organizes vector data for efficient similarity search. Each index can store vectors at scale and supports metadata filtering to refine search results.

Vector indexes are purpose-built for:

Cost-effective storage of large vector datasets

Similarity queries
Metadata filtering during queries
Integration with Amazon Bedrock Knowledge Bases and Amazon OpenSearch Service

When creating a vector index, you specify:

Distance metric (Euclidean or Cosine) for measuring similarity.
Number of dimensions for vectors.
Optional non-filterable metadata keys.

Each vector in an index consists of:

A unique vector key
The vector data
Optional metadata (up to 40 KB per vector, with 2 KB limit for filterable metadata)

Writes to vector indexes are strongly consistent, ensuring that subsequent queries include the most recently added data. You can attach metadata to vectors to filter future queries based on conditions like dates, categories, or user preferences. You can use bucket policies to grant or restrict access to specific indexes and vectors within your vector bucket.

To create a vector index in the console, find the Vector indexes section and click Create vector index.

Properties

Vector index name: The name must be 3 to 63 characters and unique within this vector bucket. Valid characters are a-z, 0-9, hyphens (-), and dots (.).
Dimension: A dimension is the number of values in a vector. A larger dimension needs more storage. A dimension must be an integer between 1 and 4096.
Distance metric: The distance metric measures the distance between a query vector and stored vectors.
- Cosine: Measures similarity between two vectors, based on directions, not magnitudes.
- Euclidean: Measures straight-line distance between two vectors, using both directions and magnitudes.

Demo Time!

This practical example is a command-line interface (CLI) application that demonstrates a complete, end-to-end RAG workflow. It showcases how to leverage the new S3 Vector Buckets as a serverless vector store eliminating the need for a separate database. The script handles the entire process, from parsing PDFs and creating vector embeddings with a local model to storing those embeddings in S3 and then querying them to generate context-aware responses. This provides a clear runnable blueprint for building a scalable and cost-effective RAG solution.

The following Python script demonstrates how to build a simple, local RAG application that leverages Amazon S3 Vectors. It uses Ollama for running the embedding and chat models locally, and the boto3 library to interact with the S3 Vectors API. This provides a powerful, scalable, and cost-effective way to build production-ready RAG applications.

Learn More About Amazon S3 Vector

Written by: Rafael Miguel

Rafael Louie Miguel, also known as Kuya Egg, is an AWS Certified Cloud Practitioner and Certified Solutions Architect Associate. An intern at Tutorials Dojo. He serves as the Director of Technology at AWS Cloud Club PUP, the first AWS Cloud Club in the Philippines. He is an undergraduate at the Polytechnic University of the Philippines currently pursuing a Bachelor’s degree in Information Technology.