# Example Docker Compose — copy to docker-compose.yml and customise.
# Build locally:  docker compose build
# Pull pre-built: docker compose pull
# Run:            docker compose up -d
services:
  backend:
    # ── Pre-built image (pull from registry) ──────────────────
    image: gitea.kansaigaijin.com/Jamie/Nat20-Notes/backend:latest
    # ── Or build locally (uncomment below) ────────────────────
    # build:
    #   context: ./backend
    #   target: runtime

    restart: unless-stopped
    environment:
      # ── Setup wizard ──────────────────────────────────────
      # "false" (or omit) to show the wizard on first run.
      # Set to "true" after onboarding completes.
      NAT20_ONBOARDING_COMPLETED: "false"

      # ── Transcription ─────────────────────────────────────
      # Model size:  tiny | base | small | medium | large-v3
      NAT20_WHISPER_MODEL: medium
      # Compute precision:  int8 (fastest/least VRAM)
      #                    | float16 (more accurate)
      #                    | float32 (full precision, slowest)
      NAT20_WHISPER_COMPUTE_TYPE: int8

      # Required for speaker diarization (accept HF gated-model terms first)
      # NAT20_HF_TOKEN: "hf_..."

      # ── LLM backend (Ollama — local) ──────────────────────
      NAT20_OLLAMA_HOST: http://localhost:11434
      NAT20_OLLAMA_MODEL: qwen2.5:7b

      # ── LLM backend (API — OpenAI-compatible) ─────────────
      # Uncomment these to use a hosted API instead of Ollama
      # NAT20_API_BASE_URL: https://api.openai.com/v1
      # NAT20_API_KEY: "sk-..."
      # NAT20_API_MODEL: gpt-4o-mini

      # ── Summarization ─────────────────────────────────────
      # Target words per chunk. Long transcripts are split into
      # chunks, each summarized separately.  Lower = more LLM
      # calls but finer granularity.   Higher = more context per
      # chunk but may exceed the model's context window.
      # Default 2500 works for most models (8K–128K context).
      # NAT20_CHUNK_WORD_TARGET: "2500"

      # Campaign context injected into every summarization prompt.
      # Inline string or path to a file inside the container.
      # NAT20_WORLD_CONTEXT: ""
      # NAT20_WORLD_CONTEXT_PATH: /data/campaign-context.txt

      # Player recap format:  story | diary | bullets | custom
      # NAT20_PLAYER_RECAP_STYLE: story
      # NAT20_PLAYER_RECAP_CUSTOM_PROMPT: ""

    volumes:
      # ── Data persistence ──────────────────────────────────
      # Option A: Named volume (auto-managed, no host path needed)
      - app_data:/data
      # Option B: Host bind mount (replace with your path)
      # - /mnt/media/dnd-sessions:/data

      # HuggingFace + Torch model caches (avoid re-downloading)
      - hf_cache:/root/.cache/huggingface
      - torch_cache:/root/.cache/torch

    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    ports:
      - "8000:8000"

  frontend:
    image: gitea.kansaigaijin.com/Jamie/Nat20-Notes/frontend:latest
    # build: ./frontend
    restart: unless-stopped
    depends_on:
      - backend
    ports:
      - "8020:8020"

volumes:
  app_data:
  hf_cache:
  torch_cache: