Question

A media company is building a RAG application using Amazon Bedrock Knowledge Bases backed by Amazon OpenSearch Serverless. The knowledge base contains 40 million document chunks that must all remain searchable, but 90% of the corpus is historical content that is queried fewer than once per day, while 10% is recent content queried thousands of times per hour. The team wants to minimize the cost of the vector store while keeping all content available for retrieval. Which approach best balances cost and availability?

Accepted Answer

Configure the OpenSearch Serverless vector search collection to store the hot 10% in memory and let the remaining 90% use tiered storage, so infrequently queried vectors are served from lower-cost storage while remaining searchable.

Answer

Split the corpus into two separate Bedrock Knowledge Bases and delete the historical knowledge base each night, re-ingesting it only when a historical query arrives.

Answer

Move the historical 90% of chunks out of the vector store into an Amazon S3 bucket and rely on S3 Select to perform semantic similarity search on demand.

Answer

Reduce the embedding dimension for the historical chunks to 128 while keeping recent chunks at 1024 dimensions within the same OpenSearch index to shrink storage.