RADAR — Deployment

Docker Compose deployment for RADAR. Pulls pre-built images from ghcr.io/fairscape/ and wires them up with an optional bundled Ollama for the chat feature.

Prerequisites

• Docker with Compose v2 (docker compose ...).

• ~6 GB free disk if you enable the LLM profile (the default qwen2.5:7b-instruct model is ~4.7 GB).

• NVIDIA GPU + drivers if you want GPU-accelerated inference with the llm profile. Run ./install-nvidia-docker.sh once on a fresh Ubuntu host to install the NVIDIA Container Toolkit, then confirm Docker can see the GPU:

  docker run --rm --gpus all nvidia/cuda:12.2.0-base-ubuntu22.04 nvidia-smi

  If that prints an nvidia-smi table, you're good. Without a GPU, remove the deploy.resources.reservations.devices block from the ollama service in docker-compose.yml — Ollama will fall back to CPU (slow but works).

Quickstart

Two options — pick one:

git clone https://github.com/fairscape/radar_deployment.git
cd radar_deployment

# Option A — full demo, with chat (needs GPU, see above).
# First run pulls Ollama model weights — 5–15 min, persists in ./ollama.
docker compose --profile llm up -d

# Option B — no chat, no Ollama.
docker compose up -d

Then open http://localhost:5173 and confirm the API:

curl http://localhost:8000/api/health
# {"status":"ok","version":"..."}

Where state lives

Four bind-mounts under this directory, all created on first up:

Path	Contents
./data/	radar.db SQLite + WAL/SHM sidecars.
./vault/	Per-user PDFs and feedback.jsonl.
./chroma/	Per-user Chroma collections (vault chunk index).
./ollama/	Ollama model weights (only with --profile llm).

Override locations by setting RADAR_HOST_{DATA,VAULT,CHROMA,OLLAMA}_DIR in .env (see .env.example).

To wipe state and start fresh:

docker compose --profile llm down
rm -rf data vault chroma          # ollama/ optional — saves a re-pull
docker compose --profile llm up -d

Logs

docker compose logs -f backend
docker compose logs -f frontend
docker compose --profile llm logs -f ollama

Verify the LLM model pulled:

docker compose --profile llm logs ollama-init
# "Pull complete." means it's done.
docker compose --profile llm exec ollama ollama list

Troubleshooting

• Chat returns 503 — start with --profile llm, confirm the model pulled (ollama list), and verify RADAR_OLLAMA_URL=http://ollama:11434 in .env (the in-compose hostname is ollama, not localhost).
• Backend can't write to ./data — bind-mounted dirs inherit the container UID. If you're running rootless Docker or hit perms issues, chown -R $(id -u):$(id -g) data vault chroma.
• Port conflicts — edit the ports: mapping in docker-compose.yml (e.g. "5174:5173"). Frontend port changes also need the image rebuilt with a matching VITE_API_BASE_URL build arg.