Want to use a vector DB but don’t want to front the hourly cost of a provisioned service? Do you want damn near infinite scalability? Is sub-second query performance good enough?

Then checkout S3 Vector Buckets designed for low latency vector indexing specifically for RAG at scale though the do say it is “ideal for workloads where queries are less frequent”. I intend to test its limits.

It works just like any other VectorDB you expect allowing you to store a vast amount of metadata.

Comparing its normal pricing to S3 isn’t completely fair but here goes:

Storage cost per month is $0.06 per GB vs Standard $0.023 per 1,000 requests. You also need to include the Vector Bucket’s logical storage of vector data, key, and metadata in those calculations.

Read requests per 1,000 are $0.055 vs standard’s $0.0004 so quite a bit more expensive.

Put and other mutators are $0.20 per GB vs the standard at $0.005.

It should be noted that this does NOT include the costs of running the Embedding Model. You will have to do that to generate the index. They of course want you to do it on Bedrock.

As of now I have NOT found terraform for this but it should be out soon.

In the end I am really excited to get deep with this tech. I have a lot of data I want to cluster and make available to my AI/ML workloads.

Let me know if you have any use cases for S3 Vector buckets you want explored.