open-webui/docker-compose.yaml

services:
  ollama:
    volumes:
      - ollama:/root/.ollama
      # Read-only mount for importing custom GGUF files (e.g. via `ollama create`)
      - /home/brian-llm/models/gguf:/gguf-import:ro
    container_name: ollama
    pull_policy: always
    tty: true
    restart: unless-stopped
    image: ollama/ollama:${OLLAMA_DOCKER_TAG-latest}
    environment:
      # Never unload a loaded model on idle -- default is 5 minutes.
      - 'OLLAMA_KEEP_ALIVE=-1'
      - 'OLLAMA_NUM_PARALLEL=4'
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    # No ports published here -- Ollama has no auth of its own, so LAN access
    # goes through the ollama-auth proxy below instead. open-webui still
    # reaches it directly over the internal docker network (see
    # OLLAMA_BASE_URL), which doesn't need the token.

  ollama-auth:
    image: nginx:alpine
    container_name: ollama-auth
    restart: unless-stopped
    depends_on:
      - ollama
    volumes:
      - ./ollama-auth/default.conf.template:/etc/nginx/templates/default.conf.template:ro
    environment:
      - 'OLLAMA_AUTH_TOKEN=${OLLAMA_AUTH_TOKEN}'
    ports:
      # Reachable from the LAN directly; restrict to the LAN subnet with the
      # host firewall (see ufw rules) as defense in depth, but the real
      # boundary is now the Bearer token check in ollama-auth/default.conf.template.
      # This is plain HTTP, so the token is readable by anything already on
      # the LAN segment that can sniff traffic -- accepted here since the
      # trust boundary is "the LAN subnet," not "the wire." Revisit with TLS
      # if that assumption ever stops holding (untrusted/guest/IoT devices
      # sharing this LAN, etc).
      - '11434:11434'

  open-webui:
    build:
      context: .
      dockerfile: Dockerfile
    image: ghcr.io/open-webui/open-webui:${WEBUI_DOCKER_TAG-main}
    container_name: open-webui
    volumes:
      - open-webui:/app/backend/data
    depends_on:
      - ollama
    ports:
      # Published on all interfaces since nginx runs on a separate LAN host
      # (192.168.50.224) and must reach this port directly. Firewalling this
      # to that host is still good hygiene (see ufw/DOCKER-USER notes) but is
      # no longer the sole security boundary -- OIDC auth below is real
      # authentication against Authentik regardless of network path.
      - '${OPEN_WEBUI_PORT-3000}:8080'
    environment:
      - 'OLLAMA_BASE_URL=http://ollama:11434'
      - 'WEBUI_SECRET_KEY='
      - 'DEFAULT_MODELS=gemma-4-e2b-mtp'
      - 'ENABLE_OAUTH_SIGNUP=true'
      # Links Authentik login to the existing local account with the same
      # email instead of rejecting it as a collision -- safe here since this
      # is a single-admin personal instance, not an open-signup multi-tenant one.
      - 'OAUTH_MERGE_ACCOUNTS_BY_EMAIL=true'
      # Fill these in after creating the OAuth2/OpenID Provider + Application
      # in Authentik (see instructions) -- OAUTH_CLIENT_ID/SECRET come from
      # that provider.
      - 'OAUTH_CLIENT_ID=${OAUTH_CLIENT_ID}'
      - 'OAUTH_CLIENT_SECRET=${OAUTH_CLIENT_SECRET}'
      - 'OPENID_PROVIDER_URL=https://auth.hetherman.cloud/application/o/ai-open-webui/.well-known/openid-configuration'
      - 'OPENID_REDIRECT_URI=https://ai.hetherman.cloud/oauth/oidc/callback'
      - 'OAUTH_SCOPES=openid email profile'
      # SSO-only: Authentik becomes the sole way to authenticate, admin
      # included. If Authentik is ever unreachable, nobody can log in until
      # it's fixed -- no local password fallback.
      - 'ENABLE_PASSWORD_AUTH=false'
      - 'ENABLE_LOGIN_FORM=false'
      - 'ENABLE_SIGNUP=false'
    extra_hosts:
      - host.docker.internal:host-gateway
    restart: unless-stopped

  # ollama-keepwarm:
  #   image: curlimages/curl:latest
  #   container_name: ollama-keepwarm
  #   restart: unless-stopped
  #   depends_on:
  #     - ollama
  #   # Pings Ollama every 60s to load the default model and refresh its
  #   # keep_alive timer -- covers first boot and re-loads it any time the
  #   # ollama container itself restarts (which flushes it from memory).
  #   # Keep this model name in sync with DEFAULT_MODELS above.
  #   entrypoint:
  #     - sh
  #     - -c
  #     - |
  #       while true; do
  #         curl -s -o /dev/null -X POST http://ollama:11434/api/generate \
  #           -d '{"model":"gemma-4-e2b","keep_alive":-1}'
  #         sleep 60
  #       done

volumes:
  ollama: {}
  open-webui: {}