Exam-style
A startup is building a Bedrock Knowledge Base over 8,000 chunks of internal documentation. Staff expect sub-second answers, traffic is roughly 200 queries a day, and the monthly budget for the whole feature is small. Which vector store fits best?
Reveal the answer
C. S3 Vectors, which charges for stored vectors and per query rather than for provisioned compute
S3 Vectors answers in under a second, is documented as best suited to infrequent query workloads, and bills for stored vectors, upload volume and query requests with no provisioned compute underneath. 8,000 chunks sits far below its ceiling of 2 billion vectors per index. A Classic vector search collection bills a minimum of two OCUs for the first collection in an account, one indexing including standby and one search including a replica, which is about USD$350 a month at the us-east-1 rate of USD$0.24 per OCU-hour, before anyone queries it. A NextGen collection group can hold minimum capacity at zero and stops charging after 10 minutes of inactivity, but restoring workers adds 10 to 30 seconds to the first request after an idle gap, and 200 queries a day leaves plenty of those gaps. Aurora pauses at 0 ACUs too, and still leaves a database cluster to operate and resume. A provisioned domain bills instance hours, not index size.
Q. Which vector store suits a small Bedrock Knowledge Base that has to answer in under a second?
A. S3 Vectors. It charges for stored vectors and per query, holds no compute open between requests, and AWS documents sub-second responses for the infrequent-query workloads it targets.
Why? The idle floor belongs to Classic OpenSearch Serverless collections: two OCUs for the first collection in an account, near USD$350 a month at USD$0.24 per OCU-hour. NextGen collection groups removed that floor by scaling to zero after 10 idle minutes, and the first request afterwards takes an extra 10 to 30 seconds while workers restart. Match the store to corpus size and query rhythm.