The problem
Healthcare support questions are both “how do I…” and “what’s happening with mine?”. A plain FAQ bot can’t answer the second kind, and an AI in healthcare must never diagnose, prescribe or leak personal data.
AI case study · RAG + tool-calling agent · Healthcare
A support agent that can actually check your appointment
Pamsy Assist answers questions for a healthcare platform’s patients, providers, clinic admins and pharmacy merchants. It knows the help docs, and it can look things up in the live system instead of just pointing people to a page.
- Role
- Backend Developer
- Models
- Gemini chat + embeddings
- Timeline
- July — August 2026
One agent, two kinds of answer
- Question + role“Is my lab order ready?” from a patient
- LangChain agentGemini decides: search the docs, call a tool, or both.
- Knowledge searchHelp docs → header-aware chunks → Gemini embeddings → Chroma
- 6 live toolsAppointments, reschedule, pharmacy stock, wallet, lab orders, tickets
Decisions worth explaining
Guardrails
Helpful, not a doctor
Medical disclaimer with no diagnosis or dosages, privacy rules aligned with NDPR, GDPR and HIPAA, and hand-off to a human for disputed transactions. Answers adapt to each of the four roles.
Fits a small server
512 MB is enough
The vector store loads only on the first knowledge question, so the API starts fast on a small host.
Tested without the API
Fake embeddings, real checks
RAG tests run against a fake embeddings class, so retrieval is verified without paying for model calls. Every request is validated with Pydantic.
Problems I hit, and how I solved them
Problem
Paying to re-read the docs
Turning the help docs into embeddings costs an API call for every chunk, every time the server starts.
Solution
Embed once, keep it on disk
The knowledge base is embedded once and saved to a local vector store. Restarts reuse it, so documents are only embedded again when they change.
Problem
Abuse and runaway bills
A public chat endpoint can be spammed, and every request costs model tokens.
Solution
Limits at the door
Chat is limited to 10 requests a minute per visitor. Messages are capped at 4,000 characters and history at 50 turns, so no single request can blow up the cost.
Problem
Not every question needs live data
Hitting the platform’s backend for every message would be slow and wasteful.
Solution
Look up real data only when asked
The agent answers “how do I…” questions from the docs, and calls one of its six live tools only for “what’s happening with mine?” questions, like an appointment or a wallet balance.
How it’s built
FastAPI · LangChain agent (LangGraph under the hood) · Gemini chat and embeddings · Chroma · typed tools calling the Laravel backend over httpx · slowapi rate limiting.