GenAI ChatBot with Amazon Bedrock Agent

Table of Contents

• GenAI ChatBot with Amazon Bedrock Agent
  • Table of Contents
  • Introduction
  • Prerequisites
  • Target technology stack
  • Architecture
  • Deployment
    • Local development setup
    • Deploy to AWS
    • Cleanup
  • Useful CDK commands
  • High-level Code Structure
  • Models
  • Customize the chatbot with your own data
    • For Knowledgebase Data Integration
      • 1. Data Preparation
      • 2. Configuration Adjustments
    • For Structural Data Integration
      • 1. Data Organization
      • 2. Configuration and Code Updates
    • General Update

Introduction

This GenAI ChatBot application is built with Amazon Bedrock, including KnowledgeBase, Agent, and additional AWS serverless GenAI solutions. The solution showcases a Chatbot that understands EC2 instances and their pricing. It demonstrates Amazon Bedrock's capabilities to convert natural language into Amazon Athena queries and to process complex data sets.

LlamaIndex is used for data processing and text-to-SQL retrieval. The solution integrates:

• Amazon S3 for storage
• Amazon Bedrock KnowledgeBase for retrieval augmented generation (RAG)
• Amazon Bedrock Agent to execute multi-step tasks across data sources
• AWS Glue to prepare data
• Amazon Athena to execute efficient queries
• AWS Lambda for serverless compute
• Amazon ECS with Fargate for the Streamlit frontend
• Amazon OpenSearch Serverless as the vector database for the KnowledgeBase

Prerequisites

• Docker
• AWS CDK Toolkit 2.240.0+, installed and configured. For more information, see Getting started with the AWS CDK.
• Python 3.13+, installed and configured. For more information, see Python Downloads.
• An active AWS account
• An AWS account bootstrapped by using AWS CDK in us-east-1 or us-west-2
• Enable the following model access in Amazon Bedrock:
  • Claude Sonnet 4.6 (Agent foundation model and text-to-SQL)
  • Claude Sonnet 4.5 (available as alternative)
  • Claude Haiku 4.5 (available as alternative)
  • Titan Embedding Text v1 (KnowledgeBase embeddings)

Target technology stack

• Amazon Bedrock (Agent, KnowledgeBase)
• Amazon OpenSearch Serverless
• Amazon ECS (Fargate)
• AWS Glue
• AWS Lambda (Python 3.13)
• Amazon S3
• Amazon Athena
• Elastic Load Balancer

Architecture

The chatbot uses two data paths:

1. Qualitative questions (e.g., "What is Amazon EC2?") → Bedrock Agent routes to KnowledgeBase → RAG over EC2 User Guide documentation stored in OpenSearch Serverless
2. Quantitative questions (e.g., "What is the cheapest EC2 instance?") → Bedrock Agent routes to Action Group → Lambda converts natural language to SQL via LlamaIndex → Athena queries EC2 pricing data

The Streamlit frontend runs on ECS Fargate behind an Application Load Balancer and invokes the Bedrock Agent through a Lambda function.

Deployment

Local development setup

Add a .env file to code/streamlit-app/ folder:

ACCOUNT_ID = <Your account ID>
AWS_REGION = <Your region>
LAMBDA_FUNCTION_NAME = invokeAgentLambda

Deploy to AWS

Create and activate a virtual environment:

python3 -m venv .venv
source .venv/bin/activate

Install dependencies:

pip install -r requirements.txt

Bootstrap (first time only):

cdk bootstrap

Synthesize and deploy:

cdk synth
cdk deploy

The first deployment takes approximately 30-45 minutes to build Docker images. Subsequent deployments take about 5-8 minutes.

Once deployed, access the chatbot via the Streamlit app URL in the CloudFormation stack outputs.

Cleanup

cdk destroy

You may also need to manually delete the S3 buckets generated by the CDK.

Useful CDK commands

• cdk ls — list all stacks in the app
• cdk synth — emit the synthesized CloudFormation template
• cdk deploy — deploy this stack to your default AWS account/region
• cdk diff — compare deployed stack with current state
• cdk docs — open CDK documentation
• cdk destroy — destroy one or more specified stacks

High-level Code Structure

.
├── app.py                            # CDK app entry point
├── cdk.json                          # CDK configuration and context
├── requirements.txt                  # CDK Python dependencies
├── assets/
│   ├── agent_api_schema/             # Bedrock Agent OpenAPI schema
│   ├── agent_prompts/                # Agent prompt templates
│   ├── data_query_data_source/       # EC2 pricing CSV data for Athena
│   ├── knowledgebase_data_source/    # EC2 User Guide docs for RAG
│   └── diagrams/                     # Architecture diagrams
└── code/
    ├── code_stack.py                 # CDK stack definition (all AWS resources)
    ├── lambdas/
    │   ├── action-lambda/            # Text-to-SQL via LlamaIndex (Docker, Python 3.13)
    │   ├── create-index-lambda/      # Creates OpenSearch Serverless index
    │   ├── invoke-lambda/            # Invokes Bedrock Agent (called by Streamlit)
    │   └── update-lambda/            # Post-deployment: Glue crawler, KB sync, agent alias
    ├── layers/
    │   ├── boto3_layer/              # Shared boto3 layer (v1.42.56)
    │   └── opensearch_layer/         # OpenSearch client layer (v3.1.0+)
    └── streamlit-app/                # Streamlit frontend (Docker, Python 3.13-slim)

Models

The application uses the following Amazon Bedrock models:

Claude Sonnet 4.5 and 4.6 are cross-region inference only — they require inference profile IDs (prefixed with us.) rather than raw model IDs.

Available model alternatives in code/lambdas/action-lambda/connections.py:

• Sonnet46 — Claude Sonnet 4.6 (default)
• Sonnet45 — Claude Sonnet 4.5
• Haiku45 — Claude Haiku 4.5

Customize the chatbot with your own data

For Knowledgebase Data Integration

1. Data Preparation

• Place your dataset in the assets/knowledgebase_data_source/ directory.

2. Configuration Adjustments

• Update cdk.json → context/config/paths/knowledgebase_file_name with your filename.
• Update cdk.json → context/config/bedrock_instructions/knowledgebase_instruction to reflect your dataset.

For Structural Data Integration

1. Data Organization

• Create a subdirectory in assets/data_query_data_source/ (e.g., tabular_data).
• Place your structured dataset (CSV, JSON, ORC, or Parquet) in the subdirectory.
• To connect to an existing database, update create_sql_engine() in code/lambdas/action-lambda/build_query_engine.py.

2. Configuration and Code Updates

• Update cdk.json → context/config/paths/athena_table_data_prefix to match your data path.
• Update code/lambdas/action-lambda/dynamic_examples.csv with text-to-SQL examples for your dataset.
• Update code/lambdas/action-lambda/prompt_templates.py to reflect your table schema.
• Update cdk.json → context/config/bedrock_instructions/action_group_description for your action lambda.
• Update assets/agent_api_schema/artifacts_schema.json to reflect your action lambda's API.

General Update

• Update cdk.json → context/config/bedrock_instructions/agent_instruction with a description of the agent's purpose for your use case.