Design a production RAG architecture on AWS with Bedrock Knowledge Bases

Ingest documents to S3, chunk and embed them into a vector store such as OpenSearch Serverless via Bedrock Knowledge Bases, retrieve with hybrid search and reranking at query time, and ground generation with citations plus Guardrails contextual grounding checks.

RAG quality is dominated by retrieval, not by the generating model. Teams that report poor RAG almost always have a retrieval problem they are trying to fix with prompt engineering.

Pipeline

``` S3 documents → Knowledge Base ingestion (chunk → embed → index) ↓ OpenSearch Serverless / Aurora pgvector ↓ Query → embed → hybrid retrieve (vector + BM25) → rerank → prompt with context ↓ Bedrock model…