I Built a RAG-Powered AI Barista Agent with Google ADK and Cloud Run

_Part of Google's Gen AI Academy APAC — Cohort 3
_
Retrieval-Augmented Generation (RAG) sounds abstract until you actually build something with it. So for Track 1 of Google's Gen AI Academy APAC, I did exactly that — deployed a RAG-powered AI agent for a fictional coffee shop, using Google's Agent Development Kit (ADK), Streamlit, and** Cloud Run**.

The Idea

Meet the "barista agent" — a chat assistant that answers customer questions about a coffee menu using real menu data, not guesses. That's the core idea behind RAG: instead of letting a model hallucinate an answer, you ground it in actual, retrievable data first.

The Stack
🧠 Google ADK — defines the agent's logic and behavior
📋 A mock menu data source — the "retrieval" part of RAG
💬 Streamlit — a clean chat UI, so it feels like a real product
☁️ Cloud Run — containerized, scalable deployment
🔐 A scoped IAM service account — least-privilege from day one

*What I Learned
*

1. RAG is mostly plumbing. The hard part isn't the "AI" — it's wiring a model to a trustworthy data source before generation happens. Get that right, and the model's job gets easy.

2. Cloud Run is great for agent prototypes. No heavy DevOps, just deploy and go — perfect for testing an idea fast.

3. Grounded agents are trustworthy agents. An assistant that pulls real data (or says "I don't know") beats one that improvises confidently, every time.

Why It Matters

This isn't just a lab exercise. A small business — a real café, with no engineering team — could genuinely use something like this: a simple assistant answering menu questions or helping staff quickly check details, without a custom app or call center. That's the whole spirit of Gen AI Academy: small, complete examples of AI solving real, local problems.