DEV Community

Cover image for I stopped talking about AI and started to build: Build a RAG System with Amazon Bedrock, Titan Embeddings & Knowledge Bases
A. Sinai Rivera
A. Sinai Rivera

Posted on

I stopped talking about AI and started to build: Build a RAG System with Amazon Bedrock, Titan Embeddings & Knowledge Bases

#aws #rag #ai #tutorial

TL;DR:
In this lab, you’ll build a complete RAG workflow using Amazon Bedrock Knowledge Bases on top of your own PDFs stored in S3. Titan Embeddings will automatically create the vector index so you can run grounded, document-aware queries with a Bedrock model. In about 25–35 minutes, you’ll deploy the full architecture in us-east-1, validate responses, review costs, monitoring, and security, and walk away with a clean blueprint to truly understand GenAI—and explain it to others.

I built this lab because I want to truly understand GenAI — not just for an exam, but to teach it, document it, and help others who are starting like me.

These topics were tricky inside sandboxes, so now I’m rebuilding everything from scratch in a real AWS account to see how things work internally.

🧭 Quick Metadata

Field Value
CB Category AI/ML
AWS Services Amazon Bedrock, Knowledge Bases, Amazon S3, Titan Embeddings
Prerequisites AWS account, Bedrock enabled, S3 permissions, region: us-east-1
Estimated Cost Under \$0.50
Architecture See diagram below

🗺️ Table of Contents

💡 Why it matters

Companies want to integrate generative AI into their apps, but raw models aren’t enough. They need answers based on internal knowledge: PDFs, manuals, policies, emails, reports.

That’s what RAG does. AWS makes this easier with Knowledge Bases for Amazon Bedrock, which handles:

  • Document extraction
  • Embedding generation (Titan)
  • Vector indexing
  • RAG orchestration
  • Grounded responses

This lab was hard for me when using sandboxes, so I rebuilt it from scratch in a AWS Free Tier, and now I’m sharing the exact steps.

🧰 Architecture / What you will build

Simple flow diagram: 


📄 Documents (PDF/Markdown/TXT)
⬇️
🪣 Amazon S3 (raw documents)
⬇️
🧠 Knowledge Base

Titan Embeddings

OpenSearch Vector Index
⬇️
🤖 Amazon Bedrock Model (Nova Micro / Claude)
⬇️
💬 Grounded, context-aware answer
Enter fullscreen mode Exit fullscreen mode

Key Notes

  • Titan automatically generates embeddings.
  • Vector store is fully managed (OpenSearch Serverless).
  • Bedrock uses ONLY your documents to answer.

✅ Prerequisites

  • AWS Free Tier account
  • Region us-east-1
  • Bedrock enabled
  • IAM permissions:
    • AmazonBedrockFullAccess
    • AmazonS3FullAccess

🛠️ Step-by-step

1) Create an S3 bucket

  1. Open AWS Console → search S3
  2. Click Create bucket
  3. Use:
  • Bucket name: 
kb-sina-rag-lab
Enter fullscreen mode Exit fullscreen mode
  • Region: us-east-1
  • Block Public Access: ALL ON
  • Encryption: default (SSE-S3)

Captura de pantalla del panel de Amazon S3 mostrando la creación de un bucket nuevo, con campos para asignar el nombre del bucket, región y configuraciones opcionales de seguridad y versionado.

Screenshot of the Amazon S3 console showing the creation of a new bucket, including fields for bucket name, AWS Region, and block public access settings.

2) Upload your documents

  1. Go to your bucket: kb-sina-rag-lab
  2. Click Upload
  3. Click Add files
  4. Choose your PDFs or TXT files
  5. Click Upload

Screenshot of the Amazon S3 console showing the bucket configuration settings, including versioning, server-side encryption (SSE-S3), and default data protection options before creating the bucket.

Screenshot of the Amazon Bedrock Knowledge Bases console showing the configuration page where you select the data source type (S3), choose an existing S3 bucket, and optionally specify a prefix for document ingestion.

Screenshot of the Amazon Bedrock Knowledge Bases console showing the configuration step to select the embedding model, with

3) Create the Knowledge Base

Go to Amazon Bedrock → Knowledge Bases → Create knowledge base

3.1 Basic settings

  • Knowledge Base Name:: Type the name of your Kb 
rag-lab-sina-kb
Enter fullscreen mode Exit fullscreen mode
  • (Optional) Description: “RAG Knowledge Base for GenAI testing.”

Screenshot of the Amazon Bedrock Knowledge Bases console showing the option to enable or disable S3 object metadata filtering, with the toggle set to “Off” before continuing to the next configuration step.

3.2 Data source → Amazon S3

Screenshot of the Amazon Bedrock Knowledge Bases console showing the vector index configuration page, including the selection of the vector database type and the option to create or use an existing index.

Select your S3 bucket.

Screenshot of the vector index creation step in Amazon Bedrock Knowledge Bases, showing the selection of OpenSearch Serverless as the engine and configuration details such as collection name, vector dimensions, and field mappings.

3.3 Embeddings model

Choose:

Titan Text Embeddings v2: 3.3 Embeddings model: Titan Text Embeddings v2. Este modelo genera los embeddings para el vector store.

Screenshot of the Amazon Bedrock Knowledge Bases permissions page, showing the IAM role auto-generated by Bedrock with access to the selected S3 bucket and OpenSearch Serverless collection, before proceeding to create the Knowledge Base.

3.4 Vector store

Choose:

OpenSearch Serverless: OpenSearch Serverless (auto-configured). And leave the automatic configuration (it’s perfect for this lab).

Screenshot of the Amazon Bedrock Knowledge Bases console showing the final review screen, summarizing the S3 data source, embedding model, vector index configuration, and IAM role before creating the Knowledge Base.

Knowledge Base created:

Screenshot of the Amazon Bedrock Knowledge Bases console after successful creation, showing a confirmation message and options to proceed with data ingestion or further configuration.

4) Sync your data

  • On the same screen, scroll to Data source
  • Click: Sync

Screenshot of the Amazon Bedrock Knowledge Bases console showing the ingestion page, where you can upload or sync documents from the configured S3 bucket to populate the Knowledge Base.

Status should show:

  • ✔️ Sync successful
  • ✔️ Status: Up to date

This may take between 30 seconds and 2 minutes depending on the size of the PDFs.

Screenshot of the Amazon Bedrock Knowledge Bases ingestion history, displaying a successful ingestion job with details such as ID, status, start time, and duration.

5) Test your RAG system

Here you will see:

  • How the Knowledge Base searches for context in your documents
  • How it selects the relevant chunks

1. On your current screen (rag-lab-sina-kb)
Look for the button: Test Knowledge Base (and click it). This will open a simple chat console.

Screenshot of the Amazon Bedrock Knowledge Bases query console showing a natural language question being asked against the indexed documents, with the model returning a grounded answer based on the ingested PDF content.

⚠️ Important: Before testing the Knowledge Base, make sure you have at least one generative model enabled in Amazon Bedrock.

To use the Test tab of the Knowledge Base, you need to select a generative model. In new AWS accounts (including Free Tier), Titan and Claude models may not be available immediately.

The compatible option that is enabled by default is: Amazon Nova Micro

Screenshot of the Amazon Bedrock Playground showing a chat interface where a natural language query is sent to the Knowledge Base, and the model returns an inline grounded response based on the ingested documents.

After enabling it, you'll be able to test your queries.

2.Test with a prompt:

Example:

What do the documents say about Amazon Bedrock?
Enter fullscreen mode Exit fullscreen mode

Screenshot from the Amazon Bedrock Playground showing a natural language query sent to a Knowledge Base, with the model returning a grounded answer along with a citations panel that references the exact PDF and page used to generate the response.

3.Verify grounding:

Screenshot of the citations panel in the Amazon Bedrock Playground, showing the source PDF used for grounding, including the file name and the highlighted page reference from which the model retrieved its answer.

Screenshot of the Amazon CloudWatch console showing log groups related to Amazon Bedrock Knowledge Bases operations, including ingestion and query logs, which can be expanded for troubleshooting and monitoring.

📊 Observability (CloudWatch)

In this lab, we work with OpenSearch Serverless, which is managed by AWS. Unlike traditional OpenSearch, this version does not create log groups automatically unless we enable specific logging policies.

In CloudWatch → Log groups, you’ll notice that no logs are generated by default. This is normal for this type of managed service and also helps keep the lab within the Free Tier.

However, we can observe lab activity in:
✔️ Bedrock Metrics

CloudWatch → Metrics → Bedrock
Here you can see model invocations and the latency of each request.

  • In the AWS Console → search for CloudWatch
  • Left menu → Metrics
  • Select Bedrock
  • Then choose All Models Lds

Screenshot of the CloudWatch Logs viewer showing detailed log events for a Bedrock Knowledge Base ingestion job, including timestamps, status updates, and processing information for troubleshooting purposes.

Screenshot of the Amazon CloudWatch Logs interface displaying the execution details of a Bedrock Knowledge Base query, including the model used, request parameters, latency metrics, and grounding information for observability and debugging.

Click on all metrics

Screenshot of the Amazon S3 console showing a PDF document stored in the bucket used as the data source for the Bedrock Knowledge Base, including file metadata such as size, last modified date, and object URL.

✔️ CloudTrail

In the console, search for CloudTrail.

  • Open the service
  • In the left menu, select Event history
  • In the filter bar, choose Event Source
  • In the search box, type: bedrock.amazonaws.com

Screenshot of the PDF opened in the Amazon S3 object preview, displaying the original source document used for the Bedrock Knowledge Base ingestion, confirming the content grounding origin.

This confirms that the Knowledge Base is working even if OpenSearch logs are not generated automatically.

🔐 Security Best Practices

Even though this is a simple lab, it’s important to follow best practices from the start. Here are the recommendations specifically for a Bedrock Knowledge Base:

  • Ever expose your bucket (keep Block Public Access ON).
  • Create separate roles if you plan to connect the Knowledge Base to applications.
  • Use plain-text documents whenever possible (better chunking).

💸 Cost Analysis

Estimated cost of the lab:

  • Titan Embeddings → a few cents for generating embeddings (very cheap).
  • Bedrock – Invocations → only a couple of calls (less than 1 cent each).
  • OpenSearch Serverless (Vector Store) → on-demand mode, the lowest possible cost.
  • Amazon S3 → very small storage (your PDFs and TXT files), practically $0.
  • Approximate total: $0.20–$0.50

🧯 Troubleshooting

Issue Cause Fix
KB doesn’t sync Wrong S3 prefix/path Re-check bucket
Missing documents Missing IAM permissions Update KB execution role
No grounding Bad PDF parsing Prefer TXT/Markdown

🚀 What’s Next

  • Connect this KB to a Bedrock Agent
  • Build a serverless RAG API using Lambda + API Gateway
  • Add authentication with Amazon Cognito

📚 Official Resources

Top comments (0)