# Single service: no queue, no cache, no state store. The pseudonym mapping is
# returned to the caller, so there is nothing to persist.
#
#   docker compose up --build -d
#   curl -s localhost:8080/health
services:
  pii-guard:
    build:
      context: .
      args:
        # Empty TORCH_INDEX switches to the CUDA wheels; see Dockerfile.
        TORCH_INDEX: https://download.pytorch.org/whl/cpu
    image: pii-guard:cpu
    container_name: pii-guard
    restart: unless-stopped
    # Bound to loopback on purpose. This endpoint accepts arbitrary text and
    # spends CPU on every request -- do not expose it to a network you do not
    # control without a proxy in front. See docs/usage.md, "Развёртывание".
    ports:
      - "127.0.0.1:8080:8080"
    environment:
      # Uncomment to require Authorization: Bearer <key>. No default on purpose.
      # - PII_GUARD_API_KEYS=${PII_GUARD_API_KEYS}
      - PII_GUARD_DEVICE=auto
      - PII_GUARD_MAX_TEXT_CHARS=100000
      - PII_GUARD_MAX_TEXTS=64
      # Weights are baked into the image; keep the runtime off the network.
      - HF_HUB_OFFLINE=1
    # No volume on /opt/hf-cache on purpose. The weights are baked into the
    # image at that path and HF_HUB_OFFLINE=1 keeps the runtime off the network,
    # so there is nothing to share and nothing to download. A named volume there
    # would instead shadow the baked weights: Docker fills such a volume only
    # while it is empty, so after rebuilding with a new NER_MODEL/NER_REVISION
    # the container would keep serving the previously cached weights and, being
    # offline, could not correct itself.