rdarrylr/aws-vector-hosting-comparison

This project is a hands-on cost and speed comparison tool for teams trying to decide where to store vector embeddings on AWS. Vector embeddings are the numeric representation of text or images that AI search and recommendation systems rely on, and there are several different places you can store them, each with its own pricing and performance tradeoffs. This repository sets up four of those options at once, Amazon S3 Vectors, Amazon OpenSearch Serverless, Aurora PostgreSQL with pgvector, and Pinecone, so you can measure them side by side instead of guessing. It loads the same 50,000 sample embeddings into all four stores, runs the same search queries against each, and records real response time measurements. It then prints an estimated monthly bill for each option at three different scales, from 10 million to 1 billion vectors, across six typical usage patterns. There is also a standalone cost calculator that works without touching AWS at all, so you can plug in your own expected traffic numbers and see projected costs before committing to a store. Everything needed to set this up is included: a Terraform configuration to create the AWS infrastructure, and Python scripts to generate the embeddings, load them, and run the benchmark. A Makefile ties the whole workflow together with simple commands like build, load data, run benchmark, and tear everything down. Pinecone is optional and can run entirely on its free tier for this demo. Running the full comparison once, at the 50,000 vector scale used in the demo, is expected to cost well under one US dollar in AWS charges. The project also sets up an automated benchmark that reruns every six hours if you leave the infrastructure running, useful for tracking performance trends over time rather than a single snapshot. This is written in Python and is companion code to a blog post explaining the same comparison in more detail.