Capstone Project · 4 Sessions
Live CMS Data · Google ADK · RAG Agent
Section 1 of 4

Session 1 — Why Healthcare AI Must Use RAG

Healthcare is one of the highest-stakes domains for AI hallucination. A chatbot that confidently states the wrong Medicare Part B premium, the wrong deductible, or the wrong enrollment deadline doesn't just give a bad user experience — it can cost a patient thousands of dollars, cause a coverage gap, or delay critical care. The figures change every year. A model's training cutoff guarantees its healthcare numbers are already out of date.

The standard fix is Retrieval-Augmented Generation: before the model answers, it searches a curated, up-to-date knowledge base and constructs its answer from retrieved passages rather than from memory. This doesn't eliminate all errors, but it makes errors detectable — because every claim in the answer is traceable to a specific retrieved chunk that the user can verify.

For this capstone, our knowledge base is built from real CMS (Centers for Medicare & Medicaid Services) public data: Medicare Parts A, B, C, and D; Medicaid; enrollment periods; preventive services; Medicare Savings Programs; and appeals. CMS is the federal agency that administers Medicare and Medicaid — its published figures are the authoritative source.

Live Demo: Medicare & Medicaid RAG Agent

Ask anything about Medicare or Medicaid. Watch the agent retrieve official CMS passages before composing its answer — never from memory alone.

Requires GEMINI_API_KEY in .env.local · Answers are informational only — not personalized advice. Always verify with Medicare.gov or 1-800-MEDICARE.

Knowledge Check

5 questions — answer all, then submit

1. Why is RAG especially critical for a healthcare information agent compared to relying on the model's training memory?

2. What is the standard 2024 Medicare Part B monthly premium for most beneficiaries?

3. How long is the Medicare Initial Enrollment Period (IEP)?

4. Why does the capstone agent have a calculate tool in addition to the RAG tool?

5. What does grounding a healthcare agent's answers in retrieved CMS passages accomplish beyond accuracy?