A patient-facing RAG chatbot that books, charges and routes

Production system for a US healthcare team. Answers grounded in their own documents with traceable retrieval, and it escalates to a person before anything clinical.

  • RAG
  • AI Product
  • AI Agents

Client work under NDA, so this describes the architecture rather than the client.

A chatbot that does things.


Built for a US healthcare team: a patient-facing assistant that answers from their own documents, and then acts on it — booking appointments, taking payments, routing to the right specialist.


Wrong answers were the risk that actually mattered, so most of the work went into retrieval rather than prompting. Hybrid search over a vector database, reranking to narrow candidates, and an answer layer that cites what it used and refuses when nothing relevant comes back.


Anything approaching clinical judgement escalates to a person. That boundary was decided before any code was written.

Stack

  • OpenAI API
  • Vector database
  • Hybrid search
  • Reranking

Engagement

Back to work

Working on something in this territory? Tell us what you’re building