monolithic-app

Marketplace Assistant - Initial Monolithic Application

This is the initial implementation of our AI-powered Marketplace Assistant. It's a monolithic application that handles all aspects of the shopping experience, from product recommendations to delivery estimation, purchase processing, and notifications.

Architecture

This a classic Agent setup where an LLM uses tool calling to find info and execute tasks.

There's a minor twist where we also call another Agent using a tool call (simulated for delivery estimates).

• Understanding user requests
• Product recommendations
• Inventory management
• Purchase processing
• Delivery estimation
• Customer notifications

We can break this into a crew of agents, but we’d still be dealing with similar limitations for production!

Design Limitations

This initial implementation has several limitations:

1. High Latency: The monolithic agent must process all aspects of a user request sequentially.
2. Token Inefficiency: The whole context is passed to the LLM for every operation, even for deterministic tasks.
3. Reliability Issues: A failure in any component impacts the entire application.
4. Debugging Complexity: Difficult to isolate issues within the monolith.
5. Scalability Challenges: The entire application needs to scale together.
6. Managing State Across failures is hard: When failures occur mid-transaction, the system can be left in an inconsistent state.

Critical Bug: Inventory Inconsistency

The application has a critical issue that demonstrates the need for proper transaction handling:

1. When a user attempts to purchase a product, the inventory is reduced BEFORE payment processing
2. If the payment fails (simulated 50% of the time) because the Payment Gateway is down, the inventory remains reduced
3. This creates a data inconsistency where products appear out of stock but were never actually purchased
4. In a production system, this would require manual intervention to fix

This bug was deliberately included to demonstrate why compensation handling is necessary in distributed systems, which we'll address in Stage 2 of the guide.

NOTE: We've kept the failure quite simple here. Payments can also be rejected quite a bit later in the future - in that case more complicated recovery is required.

Running the Application

1. Install dependencies:

npm install

2. Setup the appplication's data:

cp data.json-example data.json

3. Start the application:

npm start

4. Interact with the assistant via the terminal interface, using this interaction script.

Next Steps

In the next stage, we'll refactor this monolithic application into a multi-agent architecture using orra to orchestrate different specialised components.