Modish AI：基于 Google ADK 的多智能体零售助手

Modish AI - Multi-Agent Retail Assistant

An AI shopping assistant where 6 specialized agents work together to handle a complete retail transaction. You talk to the system in natural language, and behind the scenes, different agents handle product search, inventory checks, payments, delivery, loyalty points, and post-purchase support.

Built this to learn Google's Agent Development Kit (ADK) and understand how multi-agent orchestration works in practice. The retail domain made sense because a shopping flow naturally breaks into distinct responsibilities.

The agents

Sales Agent (orchestrator)
├── Recommendation Agent  → Suggests products based on what you ask for
├── Inventory Agent       → Checks stock availability and pricing
├── Payment Agent         → Handles checkout and payment processing
├── Fulfillment Agent     → Manages shipping and delivery tracking
├── Loyalty Agent         → Tracks reward points, applies discounts
└── Post-Purchase Agent   → Returns, exchanges, order status queries

The Sales Agent is the one you talk to. It figures out which worker agent to route your request to, collects the response, and gives you a unified answer. If your request touches multiple agents (e.g., "buy this red kurta and apply my loyalty points"), it orchestrates between them.

Example conversation

You: "Show me cotton kurtas under 2000 rupees"
→ Recommendation Agent searches the product catalog
→ Inventory Agent checks stock for each suggestion
→ Sales Agent combines both and responds with available options

You: "Buy the blue one, use my reward points"
→ Inventory Agent reserves the item
→ Loyalty Agent calculates point redemption
→ Payment Agent processes the adjusted total
→ Fulfillment Agent creates a shipping order

How I built it

The agents use Google ADK (Agent Development Kit) with Gemini as the underlying LLM. Each agent has:

A system prompt defining its role and tools
Python functions it can call (database queries, calculations, etc.)
Clear boundaries on what it should and shouldn't handle

The product catalog uses an Indian retail dataset - kurtas, sarees, electronics, etc. with prices in INR.

Tech

Agent framework: Google ADK with InMemoryRunner
LLM: Google Gemini API
Backend: Python, Flask (api_server.py for the REST API)
Frontend: React app (ey-frontend/)
Database: SQLite with product catalog and order data
Deployment configs: Railway, Render, Procfile

Running it

# install dependencies
pip install -r requirements.txt

# set up env (needs a Google Gemini API key)
cp .env.example .env
# edit .env with your GOOGLE_API_KEY

# run the CLI version
python app.py

# or run the API server + frontend
python api_server.py
cd ey-frontend && npm install && npm start

What's interesting about multi-agent systems

The tricky part isn't building individual agents - it's the orchestration. When should the sales agent delegate vs answer directly? How do you handle errors when one agent fails mid-transaction? What if two agents need the same data?

I ran into all of these. The payment agent failing after inventory was reserved was the hardest edge case - you need rollback logic that works across agents.