RAG-Based Support Automation in Healthcare Training

The problem

The institution's support team was buried under enrolment, payment, exam, certificate, and technical questions from thousands of students. Scale was not the hard part. Answering correctly required knowing the contents of 21 medical textbooks and all of the institution's procedures, and in a healthcare context a wrong answer is not an acceptable error — which ruled out a standard chatbot from the outset.

How we approached it

All 21 medical textbooks and the institution's full content went into a vector database. Every incoming question is classified first — five main categories, fourteen subcategories, and over thirty tags — then semantic search finds the most relevant source, and the answer is generated from that source alone rather than from the model's general knowledge. Confidence scoring is the backbone of the architecture: below the threshold the system does not answer at all, it routes the case to a human. In a healthcare context the right design decision was not to widen coverage but to stay silent where certainty runs out.

Outcome

• End-to-end flow from HubSpot ticket through classification and semantic search to answer and CRM update
• Automatic language detection across German and English
• Context from earlier correspondence preserved across ongoing tickets
• Answers grounded in retrieved sources rather than the model's general knowledge alone
• Cases below the confidence threshold routed to a human automatically