📚 Amazon Bedrock – Setting up RAG & Knowledge Base (Hands-on)

This guide explains how to set up a Retrieval-Augmented Generation (RAG) pipeline and a Knowledge Base in Amazon Bedrock, using Amazon S3 for storage and Amazon OpenSearch Serverless as the vector database.

1. 🔍 Prerequisites

  • IAM User (not root user)
  • Administrator Access policy for the IAM user
  • AWS services:
    • Amazon Bedrock
    • Amazon S3
    • Amazon OpenSearch Serverless (or external vector DB)
  • PDF or text document to upload (e.g., evolution_of_the_internet_detailed.pdf)

2. 🛠 Step-by-Step Setup

Step 1 – Create an IAM User

  1. Go to IAM Console → Users → Create User.

  2. Enter a username (e.g., stephane).

  3. Enable AWS Management Console Access.

  4. Set a custom password.

  5. Attach the AdministratorAccess policy.

  6. Save the sign-in URL, username, and password.

  7. Log in as the IAM user (not root).

Step 2 – Create a Knowledge Base in Amazon Bedrock

  1. In Amazon Bedrock, go to Knowledge Bases → Create Knowledge Base.
  2. Set the name (default is fine).

  1. IAM permissions → Create and use a new service role.
  2. Data Source → Select Amazon S3.
  3. Alternative sources (optional):
    • Web crawler (webpages)
    • Confluence
    • Salesforce
    • SharePoint

Step 3 – Create an Amazon S3 Bucket & Upload Documents

  1. Go to Amazon S3 → Create bucket.
    • Region: us-east-1
    • Bucket name: must be globally unique (e.g., my-demo-bucket-knowledgebase-danny)
  2. Upload your document:
    • Example: evolution_of_the_internet_detailed.pdf
  3. Confirm the object appears in the bucket.

Step 4 – Connect S3 to Bedrock Knowledge Base

  1. In Bedrock KB creation:
    • Select your S3 bucket as the data source.
    • Click Next.

  1. Embedding Model:
    • Select Amazon Titan Text Embeddings V2 (default dimensions).
  2. Vector Database:
    • For AWS exam → Amazon OpenSearch Serverless is the common choice.
    • External free option → Pinecone (free tier available).
  3. Complete the KB creation.

⚠️ Cost Warning:
Amazon OpenSearch Serverless minimum cost is ~$172/month (2 OCUs at $0.24/hour).
Delete resources after use to avoid charges.

Step 5 – Sync Data to Vector Database

  1. Open your Knowledge Base.
  2. Click Sync to push S3 data → embeddings → vector database.
  3. In OpenSearch Service:
    • View your collection and indexes.
    • Each chunk of your document is stored as a vector.

Step 6 – Test the Knowledge Base

  1. Configure a model (e.g., Anthropic Claude Haiku).
  2. Ask a question (e.g., "Who invented the World Wide Web?").
  3. Bedrock will:
    • Perform vector similarity search (KNN search).
    • Retrieve relevant chunks from the KB.
    • Augment the prompt with retrieved text.
    • Generate an answer with source citations.
  4. Click the source link → View the PDF in S3.

3. 🧠 How It Works Internally

📈 RAG Data Flow Diagram

1
2
3
4
5
6
7
flowchart TD
    A[📂 Amazon S3 PDF] --> B[✂️ Chunking & Embedding Creation<br/>(Amazon Titan)]
    B --> C[🗄 Vector Database<br/>OpenSearch Serverless]
    C --> D[🔍 KNN Similarity Search]
    D --> E[📑 Relevant Chunks Retrieved]
    E --> F[📝 Combined with Original Query<br/>→ Augmented Prompt]
    F --> G[🤖 Foundation Model Generates Answer]
  • Chunking: Splits the document into smaller parts.
  • Embeddings: Numeric vector representation of text.
  • KNN Search: Finds the k most semantically similar chunks.
  • Augmented Prompt: Original query + retrieved text → better answer.

4. 🛑 Cleanup (Avoid Unnecessary Costs)

After testing:

  1. Delete Knowledge Base in Bedrock.
  2. Delete OpenSearch Serverless collection.
  3. (Optional) Keep S3 bucket (low cost) or delete it.

5. 📌 Exam Tips

  • Always use IAM user (root user cannot create Bedrock KB).
  • Vector DB Options in AWS:
    • OpenSearch (real-time search, KNN)
    • Aurora PostgreSQL (pgvector)
    • Neptune Analytics (graph-based RAG)
    • S3 Vectors (low cost, sub-second search)
  • External: Pinecone, Redis, MongoDB Atlas Vector Search.
  • Bedrock KB supports multiple data sources, not just S3.
  • Remember: RAG = Retrieve external data + Augment prompt + Generate answer.

✅ Summary:
You’ve created a Bedrock Knowledge Base with Amazon S3 + OpenSearch, generated embeddings with Titan, performed KNN search, and tested retrieval-augmented responses with citations.