RAG on AWS with Bedrock Knowledge Bases
The RAG building block on AWS. A knowledge base ingests your documents, parses, chunks and embeds them, then returns the right passages to your LLM or agent. Bedrock can run the whole pipeline for you, or you run it on the vector store you choose.
Documents in, the right passages out. Two ways to run the store.
Customer-managed vector stores: Amazon OpenSearch (Serverless or managed clusters), Aurora PostgreSQL, S3 Vectors, Neptune Analytics, Pinecone, Redis Enterprise Cloud, MongoDB Atlas. Managed KB: seven native connectors (S3, SharePoint, Confluence, Web Crawler, Google Drive, OneDrive, custom), built-in parsing, embedding and reranking, plus Retrieve, agentic retrieval and AgentCore Gateway. Customer-managed: S3 and custom connectors. Feature and model support differ by Region.
Get passages, get an answer, or let it search in steps.
Retrieve
Retrieve → your Converse call
Returns chunks, metadata and scores. You generate with your own LLM call, so you control citations, refusals when no source answers, and guardrails.
RetrieveAndGenerate
RetrieveAndGenerate
Retrieves and generates in one call with the model you pick. Faster to start; the prompt template and the answer format are less yours to shape.
Agentic retrieval
AgenticRetrieveStream
Managed KB only. Splits the question, searches, checks whether the context is enough, searches again, then hands a synthesised context to your LLM.
Three decisions set the quality and the bill.
Managed or customer-managed
Managed bundles the store, parsing, embedding, reranking and connectors. Customer-managed lets you choose the store (OpenSearch, Aurora PostgreSQL, S3 Vectors, Neptune Analytics, Pinecone, Redis, MongoDB Atlas), the parser and the chunking.
How documents are parsed and cut
Parsing decides what happens to tables, columns and scans. Chunking cannot change once the data source exists; hierarchical chunking matches on small chunks and returns the larger parent for context.
How the right passage wins
A metadata file per document (product, version, status) lets you filter at query time. A reranker rescores the chunks so fewer reach the LLM. Agentic retrieval costs more per call; route only multi-step questions to it.
Tag values describe mechanisms, not measured figures.
A knowledge base bill, two shapes.
End-to-end agentic AI use cases
From business need to production, customer-facing included, with evaluation and ops to match.
Read the approach →The one-page RAG on AWS with Bedrock Knowledge Bases field guide.
Best practices, cost traps and performance checks, in the order you should run them. Printable, no fluff.