🧠 AgenticBI — Unified LLM-Driven E-Commerce Automation

A full agentic architecture powered by Ballerina, WSO2 Integrator, WSO2 API Manager.

This project implements a complete multi-agent generative system for handling:

• Placing Orders
• Updating Orders
• Cancelling + Refunding Orders
• Routing Messages Across Specialized Agents
• Safely calling backend tools using MCP (Model Context Protocol)
• Storing and retrieving orders through a PostgreSQL microservice
• Exposing all functionality through API Manager and an MCP Server

It includes:

• /agents → 4 LLM Agents (Router, PlaceOrder, UpdateOrder, CancelRefund)
• /backend → MCP Tool Runtime (Product Search, Order DB, Invoice Creation, Refunds…)
• /facade → WSO2 MI deployment that exposes backend tools
• /docker-compose.yml → Infra stack: PostgreSQL, Prometheus, Grafana, Jaeger
• /init.sql → DB initialization schema

🏗 Folder Structure

AgenticBI/
├── agents/          # All LLM agents + tooling integration
├── backend/         # Ballerina MCP Tools service (DB, invoices, products)
├── facade/          # WSO2 Micro Integrator deployment
├── monitoring/      # Prometheus configuration
├── docker-compose.yml
├── start_all.sh     # Script to start the entire environment
├── stop_all.sh      # Script to stop the entire environment
├── init.sql         # DB bootstrap schema for Postgres TMPFS instance
└── README.md

🔧 Requirements

Before running anything, ensure you have:

📁 Configuration Required Before First Run

1. Edit /agents/Config.toml

This file contains your OpenAI credentials, backend endpoints, tenant IDs, etc.

YOU MUST UPDATE/INCLUDE THESE FIELDS:

MCP_ENVIRONMENT = "SANDBOX"

MCP_SANDBOX_TOKEN_URL = "" MCP_SANDBOX_CLIENT_ID = "" MCP_SANDBOX_CLIENT_SECRET = ""

For now, mirror sandbox URLs for prod; in a real setup this will be a different host.

MCP_PROD_TOKEN_URL = "" MCP_PROD_CLIENT_ID = "" MCP_PROD_CLIENT_SECRET = ""

Optional; leave empty if a scope is not required

MCP_OAUTH_SCOPE = ""

MCP_SERVER_URL = "" OPENAI_API_KEY=""

[ballerina.observe]

metricsEnabled = true

metricsReporter = "prometheus"

tracingEnabled = true

tracingProvider = "jaeger"

[ballerinax.prometheus]

port = 9797

host = "0.0.0.0"

[ballerinax.jaeger]

agentHostname = "jaeger"

agentPort = 4317

samplerType = "const"

samplerParam = 1.0

reporterFlushInterval = 2000

reporterBufferSize = 1000

If this is not configured correctly:

• Agents will fail to run
• Tools won’t connect
• Router will return errors
• LLM calls may not execute due to missing credentials

2. Expose Backend Tools via WSO2 Micro Integrator (Facade)

The project includes fully prepared WSO2 MI artifacts under:

facade/src/main/java/wso2mi/artifacts

You have two supported ways to expose the backend tools from Micro Integrator to API Manager:

Option A — Expose as a Service Catalogue API (Recommended for MI ↔ APIM integration) Option B — Import the OpenAPI Definition Manually into APIM

3. Publish the MI API as an MCP Server in API Manager

✔ Step-by-step:

4.1 Deploy MI API into APIM

In API Manager Publisher:

1. Create MCP Server → point to your API
2. Configure
3. Publish the MCP Server.

🚀 Starting the Entire System

Enter start-all.sh and adjust the folder paths to your local configuration

One command:

./start_all.sh

This launches:

Check logs inside:

logs/

🛑 Stopping Everything

./stop_all.sh

This gracefully terminates:

• Agents
• Backend
• MI
• API Manager
• Docker services

🧪 Testing the System

Before running any cURL tests, you must first expose the unified Agent API (from /agents) as an AI API in WSO2 API Manager.

🔁 Step 0 — Publish the Unified Agent API as an AI API in APIM

(Optional but recommended) Apply policies:

AI governance / prompt protection

Custom mediation policies

Save → Publish the API.

Go to DevPortal, subscribe, generate tokens, and test the API.

🔍 Observability

Prometheus

http://localhost:9090

Grafana

http://localhost:3000
# Default login:
admin / admin

Jaeger (Distributed Tracing)

http://localhost:16686

🛠 Known Constraints

• The database runs on tmpfs, meaning all orders reset when Postgres restarts.
• You MUST expose the MI backend API through APIM before the agents can call tools.
• Config.toml must be updated before first run.
• Sessions stick across messages; use "new subject" to reset context routing.