Multi-Agent Post-Sale Support System

Cloud Architecture & Generative AI demonstration project.

An AI assistant designed to autonomously handle customer inquiries, inventory verification, and warranty policy lookups. The system implements a RAG (Retrieval-Augmented Generation) architecture and a task-delegation workflow built on specialized, collaborating agents.

---

System Architecture

The project is built on Amazon Bedrock and follows a task-delegation pattern across specialized agents:

1. Supervisor Agent (Post-Sale Assistant): The main orchestrator. It receives the user's query and decides whether to consult the knowledge base directly or delegate the task to the inventory agent.

2. Inventory Agent: A specialized agent for order and stock management, connected to a database through a Lambda function.

3. Knowledge Base (RAG): A knowledge store connected to a vector index in Amazon OpenSearch Serverless, enabling responses grounded in technical documents.

4. Action Handler (Lambda): Python execution logic for interacting with Amazon DynamoDB.

---

Project Structure

.
├── data/                   # Technical documentation and policies (RAG dataset)
│   └── politicas_technova_2026.pdf
├── infrastructure/         # Infrastructure as Code (IaC)
│   └── template.yaml       # AWS CloudFormation Template
├── src/                    # Business logic
│   └── index.py            # Lambda Function (Action Handler)
├── README.md               # Project documentation
└── pyproject.toml          # Python environment and dependency management

---

Tech Stack

Component	Technology
Generative AI	Amazon Bedrock — Claude 3.5 Sonnet / Haiku
Vector Database	Amazon OpenSearch Serverless
Compute	AWS Lambda (Python 3.12)
NoSQL Database	Amazon DynamoDB
Infrastructure as Code	AWS CloudFormation
Environment Management	uv (package manager and lock files)

---

Installation & Deployment

Prerequisites

• AWS account with Amazon Bedrock models enabled.
• AWS CLI configured with appropriate credentials.
• Policy document uploaded to the corresponding S3 bucket.

Deployment Steps

Some resources (Bedrock Agents, Knowledge Base, OpenSearch Serverless collection) must be created manually in the AWS console before deployment, as CloudFormation cannot provision them from scratch. Once created, provide their IDs as parameters.

aws cloudformation deploy \
  --template-file infrastructure/template.yaml \
  --stack-name technova-ai-stack \
  --parameter-overrides \
      KnowledgeBaseId=YOUR_KB_ID \
      SupervisorAgentId=YOUR_AGENT_ID \
      OSSCollectionId=YOUR_OPENSEARCH_COLLECTION_ID

---

System Capabilities

• Warranty Lookups: Precise information extraction from PDF documents stored in the Knowledge Base.
• Order Management: Read and write access to DynamoDB through Bedrock Action Groups.
• Autonomous Orchestration: The supervisor agent selects the appropriate tool based on the intent detected in the user's query.

---

Security & Governance

This project follows the Principle of Least Privilege (PoLP):

• IAM Roles: Each component (Agents, Knowledge Base, Lambda) has a dedicated role with permissions scoped exclusively to the resources it needs.
• OpenSearch Access: Data access policies are restricted to the ARNs of the Bedrock service execution roles.
• Cost Optimization: DynamoDB is configured in On-Demand mode to avoid fixed provisioning costs.