DocGemma

Source repositories: docgemma-connect (FastAPI backend) | docgemma-frontend (Vue 3 UI)

Competition: The MedGemma Impact Challenge on Kaggle

Agentic medical AI assistant powered by MedGemma, with autonomous tool calling for clinical decision support. Designed for resource-limited healthcare environments. Compatible with MedGemma 27B and MedGemma 1.5 4B.

DocGemma combines a Vue 3 web interface with a FastAPI/LangGraph agent backend that can query drug safety databases, search medical literature, manage electronic health records, and analyze medical images — all with human-in-the-loop approval for safety-critical actions.

Try DocGemma instantly on Google Colab — no local setup required. The notebook provisions an A100 GPU, deploys the full stack (vLLM + backend + frontend), and generates a public URL. Just provide a HuggingFace token with MedGemma access.

• Docker and Docker Compose v2+, or Podman with podman compose

For local GPU inference (optional):

• NVIDIA GPU — 48 GB+ VRAM for MedGemma 27B (e.g., A100, A6000) or 8 GB+ for MedGemma 1.5 4B
• NVIDIA Container Toolkit
• A HuggingFace account with access to MedGemma

For remote inference:

• A running vLLM-compatible endpoint (e.g., RunPod, Together AI)

For GPU cloud instances (RunPod, Vast.ai, Lambda, Paperspace, AWS/GCP/Azure) or bare metal with an NVIDIA GPU. Installs everything directly on the host — no containers needed.

git clone https://github.com/galinilin/docgemma-app.git
HF_TOKEN=hf_your_token_here bash docgemma-app/run-native.sh

The script auto-detects GPU VRAM (falls back to MedGemma 4B if < 40 GB), installs dependencies (Node.js, UV, vLLM), builds the frontend, and starts vLLM + the app on a single port.

git clone https://github.com/galinilin/docgemma-app.git
cd docgemma-app
cp .env.example .env

Edit .env and set your endpoint credentials:

DOCGEMMA_ENDPOINT=https://your-vllm-endpoint.example.com
DOCGEMMA_API_KEY=your-api-key-here

docker compose --profile remote up

git clone https://github.com/galinilin/docgemma-app.git
cd docgemma-app
cp .env.example .env

Edit .env and set your HuggingFace token:

HF_TOKEN=hf_your_token_here

docker compose --profile gpu up

Note: The first run downloads model weights (~54 GB for MedGemma 27B, ~8 GB for MedGemma 1.5 4B). The vLLM health check allows up to 10 minutes for the model to load.

Both profiles work with Podman. For the remote profile, use the default compose file:

podman compose --profile remote up

For the gpu profile, use the Podman-specific compose file (uses CDI instead of Docker's deploy.resources for GPU passthrough):

podman compose -f docker-compose.podman.yml --profile gpu up

Requires NVIDIA Container Toolkit with CDI configured for Podman.

Open http://localhost:8080 in your browser.

Browser ──→ Vue 3 SPA ──→ FastAPI + LangGraph Agent ──→ vLLM / MedGemma
                │                    │
                │                    ├── Drug safety (OpenFDA)
                │                    ├── Drug interactions (RxNav)
                │                    ├── Medical literature (PubMed)
                │                    ├── Clinical trials (ClinicalTrials.gov)
                │                    ├── FHIR R4 EHR (local store)
                │                    └── Medical image analysis (vision API)
                │
                └── Patient management, imaging upload, clinical notes

The Docker image bundles everything into a single container:

• Frontend: Vue 3 + TypeScript + Tailwind CSS, built and served as static files
• Backend: Python/FastAPI with a 7-node LangGraph agent workflow
• Model: MedGemma 27B or 1.5 4B via vLLM (OpenAI-compatible API)

Patient records (FHIR R4) and chat session history are stored in a Docker volume (docgemma-data). On first run, sample patient records are seeded automatically.

To reset all data to defaults:

docker compose down -v

Pin the backend and frontend to specific branches or tags:

docker compose build \
  --build-arg BACKEND_REF=v1.0.0 \
  --build-arg FRONTEND_REF=v1.0.0

Available build args:

For development, clone the source repositories directly:

• Backend: galinilin/docgemma-connect
• Frontend: galinilin/docgemma-frontend

vLLM takes a long time to start First run downloads model weights (~54 GB for MedGemma 27B, ~8 GB for MedGemma 1.5 4B). Subsequent runs use the cached weights in the vllm-cache Docker volume.

GPU out of memory Reduce VLLM_GPU_UTIL (e.g., 0.80) or VLLM_MAX_MODEL_LEN (e.g., 4096) in your .env.

Connection refused on remote profile Verify DOCGEMMA_ENDPOINT is reachable and includes the correct port (e.g., https://host:8000).

Rebuilding after source updates Force a fresh build to pull latest source code:

docker compose build --no-cache

DocGemma is a research and educational tool. It is not certified for clinical use. Do not use for real patient care decisions.