Question

A financial services company already maintains millions of pre-computed embeddings in an Amazon OpenSearch Serverless collection, generated by a legacy pipeline using a specific 1536-dimension embedding model. They now want to adopt Amazon Bedrock Knowledge Bases to power a RAG chatbot, reusing their existing vector index rather than re-embedding all documents. Which configuration approach allows the Knowledge Base to work correctly with this existing vector store?

Accepted Answer

Create the Knowledge Base pointing to the existing OpenSearch Serverless collection, and configure the vector index field mapping (vector field, text field, metadata field) along with an embedding model whose output dimension matches the stored 1536-dimension vectors.

Answer

Create the Knowledge Base with a new Amazon Aurora PostgreSQL vector store and let Bedrock automatically copy the existing embeddings from OpenSearch during the first ingestion sync.

Answer

Point the Knowledge Base to the existing collection but let Bedrock choose any embedding model, because the Knowledge Base re-normalizes stored vectors to the model's native dimension at query time.

Answer

Delete the existing index and allow Bedrock Knowledge Bases to fully re-ingest and re-embed all source documents, since reusing externally generated embeddings is not supported in any configuration.