Knowledge Bases for Bedrock — Managed RAG

Retrieval-Augmented Generation in plain terms

RAG solves a fundamental problem: language models know a lot of general knowledge, but they don't know your documents — your policies, your product manuals, your internal research.

RAG works in two stages:

Retrieval: When a user asks a question, the system searches your document store for the most relevant chunks of text.
Augmentation: Those chunks are included in the prompt to the LLM as context. The LLM answers the question using your documents as the source of truth.

Bedrock Knowledge Bases handles both stages for you — the embedding, the vector search, the retrieval pipeline. You just point it at an S3 bucket and ask questions.

Set up the Knowledge Base in the console

Upload 3-5 PDF or text documents to a new S3 bucket. Then:

In Bedrock → Knowledge bases → Create knowledge base.
Name it my-docs-kb.
Data source: S3. Point to your bucket.
Embeddings model: Amazon Titan Embeddings G1 - Text (free tier).
Vector store: Amazon OpenSearch Serverless (auto-created).

Click Create. Wait 2-3 minutes for provisioning. Then click Sync to ingest your documents. Syncing takes 1-5 minutes depending on document size.

Query the Knowledge Base programmatically

import boto3

agent_runtime = boto3.client("bedrock-agent-runtime", region_name="us-east-1")

KNOWLEDGE_BASE_ID = "YOUR_KB_ID"  # from the console
MODEL_ARN = "arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-opus-4-5-20251101-v1:0"

def query_docs(question):
    response = agent_runtime.retrieve_and_generate(
        input={"text": question},
        retrieveAndGenerateConfiguration={
            "type": "KNOWLEDGE_BASE",
            "knowledgeBaseConfiguration": {
                "knowledgeBaseId": KNOWLEDGE_BASE_ID,
                "modelArn": MODEL_ARN
            }
        }
    )
    answer = response["output"]["text"]
    citations = response["citations"]

    print(f"Answer: {answer}\n")
    print("Sources:")
    for c in citations:
        for ref in c.get("retrievedReferences", []):
            loc = ref["location"]["s3Location"]["uri"]
            print(f"  - {loc}")

query_docs("What are the main topics covered in these documents?")

Replace YOUR_KB_ID with the Knowledge Base ID from the Bedrock console (it looks like ABCDE12345).