Aurora pgvector caps Bedrock Knowledge Bases at 2,000 dimensions in single precision

A query in this design pays for two steps before the model sees anything. The embedding call turns the question into a vector, then the similarity search returns the closest chunks, typically the top five ^[4][5]. Retrieval is the half the backend controls. AWS describes OpenSearch Service as providing high-speed results from data held in memory, and S3 Vectors as providing sub-second query performance for similarity searches ^[6][9]. Those two phrases describe different things. Sub-second includes 900 milliseconds, and in-memory tells you where the data sits, which leaves the p99 an open question.

The specification most likely to disqualify an option outright is in the Aurora line. Aurora PostgreSQL with pgvector handles vectors up to 2,000 dimensions in single precision, alongside IVFFlat and HNSW indexes and L2, cosine and inner-product distance ^[8]. The ceiling binds the embedding model. If the model you want emits more than 2,000 dimensions, you either reduce the output or pick a different backend. Reducing the output changes retrieval quality, so the evaluation set has to be run again.

The 90 percent storage saving AWS attributes to S3 Vectors is stated against "traditional vector databases" ^[9]. Three things have to be true for it to reach your invoice. Storage has to dominate your vector spend ahead of query volume, your baseline has to be a provisioned cluster sized for the same corpus and kept warm, and your latency budget has to tolerate a sub-second ceiling. If query volume dominates instead, the saving applies to the smaller line.

Hybrid search is the other discriminator. In this comparison AWS credits OpenSearch with k-NN search and with hybrid search combining lexical and vector approaches ^[6]. Catalog search is the use case that wants it, because a shopper typing an exact part number is doing keyword retrieval. AWS opens its use-case walkthrough with ecommerce search that has to scale to thousands of concurrent queries during peak shopping periods ^[12].

Count deployment targets and there are four. Bedrock Knowledge Bases supports OpenSearch Managed Clusters and OpenSearch Serverless as separate backends, plus Aurora and S3 Vectors ^[7][14]. The customer-managed path also reaches three of the six services in the AWS vector portfolio ^[10][13].

AWS says the choice impacts performance and cost ^[3]. The post gives specifications, and the decision gets made on specifications, because the store holds content that was pre-processed, chunked and embedded before any query arrives ^[4]. Changing backends later means re-chunking and re-embedding the corpus into the new one.