Afya Gemma: Two-Stage Clinical RAG

Clinical decision support grounded on Kenyan MoH guidelines · Aga Khan University Hospital

Afya Gemma: Two-Stage Clinical RAG Clinical decision support grounded on Kenyan MoH guidelines · Aga Khan University Hospital GCP · GKE · Vertex AI · Terraform · Monitoring Clinician · residents & interns, AKU Hospital · Architecture component Clinician residents & interns, AKU Hospital Stage 1 · Gemini 2.5 Flash · intent + scope classifier · GCP · GKE · Vertex AI · Terraform · Monitoring · rejects out-of-scope Stage 1 · Gemini 2.5 Flash intent + scope classifier rejects out-of-scope Retrieval · ChromaDB · MoH guideline embeddings · GCP · GKE · Vertex AI · Terraform · Monitoring · RBAC-filtered Retrieval · ChromaDB MoH guideline embeddings RBAC-filtered Stage 2 · MedGemma · grounded generation · GCP · GKE · Vertex AI · Terraform · Monitoring Stage 2 · MedGemma grounded generation MoH guidelines · curated + embedded corpus · GCP · GKE · Vertex AI · Terraform · Monitoring MoH guidelines curated + embedded corpus Audit log · every query & response · GCP · GKE · Vertex AI · Terraform · Monitoring Audit log every query & response clinical question in-scope + domain guideline context embeddings log grounded answer + guideline refs graceful refusal Legend External Backend Database Security

Why two stages

  • • Flash classifier is an order of magnitude cheaper than the generator
  • • Out-of-scope questions rejected before retrieval or generation spend
  • • Domain-scoped retrieval improves MedGemma grounding

Safety & governance

  • • Every answer traceable to MoH guidelines
  • • Retrieval filtered by clinician role (RBAC)
  • • Every query and response audited

Production

  • • <300ms p95 on the retrieval path
  • • 99.9%+ uptime, hospital-platform SLA
  • • Terraform IaC, Kubernetes serving

Architecture diagram • Built with Archify • Create yours ↗ • Hover to trace • R route • Click to focus • +/− zoom • M radar • [/] views • P play story • T theme • E export