StreamDocs, a legal-document SaaS, is building an Amazon Bedrock knowledge base from a large S3 bucket of PDFs. They want vector search that supports semantic retrieval and stores document metadata (document_id, page). Which configuration best balances retrieval quality, maintainability, and AWS managed components?

The correct option leverages Amazon Bedrock's knowledge base feature, which is designed for this exact use case. It automatically handles text extraction, chunking (with recommended overlapping chunks for better context preservation), embedding generation, and vector storage, integrating seamlessly with Amazon OpenSearch Service for efficient vector search using k-NN and HNSW. Storing metadata directly in OpenSearch alongside vectors allows for effective filtering and retrieval. KMS encryption and IAM roles ensure security and controlled access. The incorrect options are flawed: Extremely small chunks (5-10 words) lose context and degrade semantic retrieval quality. DynamoDB is not optimized for vector similarity search. S3 Select is for querying object content, not semantic similarity on embeddings, and an S3 index doesn't provide vector search capabilities. Computing embeddings on-demand for every query is inefficient and costly, especially for a large dataset, and Bedrock's knowledge base is designed for persistent storage and efficient retrieval.