One signal

AWS's machine learning blog has published a selector for the customer-managed vector stores behind Amazon Bedrock Knowledge Bases, weighing Amazon OpenSearch Service, Amazon Aurora PostgreSQL with pgvector and Amazon S3 Vectors across different RAG use cases. The budget option's pitch is the eye-catcher: per the post, S3 Vectors delivers sub-second similarity search while cutting vector storage costs by up to 90 per cent against traditional vector databases.

Why it matters

The first worked example, product catalogue search, lands on OpenSearch Serverless, which pairs semantic vectors with keyword matching in hybrid search and serves faceted filtering in low milliseconds. The benchmarking underneath is real rather than decorative: seven index configurations across embedding sizes and data types, run against Amazon's ESCI shopping dataset of roughly 1.2 million products and scored on retrieval quality, latency and index size. Binary embeddings and a 32× on-disk mode shrink the index, at the price of latency.

Discuss: S3 Vectors promises up to 90 per cent off vector storage, yet the guide sends search-heavy workloads to pricier OpenSearch for hybrid retrieval: at what point does cheap storage beat keyword-plus-vector search in a RAG stack?

Independent WittyWires Watcher; not an official account or feed.