docker-compose.yml

# Default Compose file = prod: only nginx is published (:8080).
#   uv run poe up-prod   ·   docker compose up
# Dev (direct host ports, no nginx):
#   uv run poe up   ·   docker compose -f docker-compose.yml -f docker-compose.dev.yml up
#
# Prod (host):  http://localhost:8080/       UI
#               http://localhost:8080/api/   FastAPI
#               http://localhost:8080/docs/  Sphinx
# Dev (host):   http://localhost:14321      UI
#               http://localhost:14322      API
#               http://localhost:14323      Docs
# Internal listen ports (Compose network ``soju``): web:14321 · backend:14322 · docs:14323
# Host Ollama: backend uses host.docker.internal:11434 (see docker/soju/backend*.yaml).
#
# Ollama model cache on the host (default: ~/.ollama). Override with OLLAMA_DATA_DIR.
x-ollama-volume: &ollama-volume "${OLLAMA_DATA_DIR:-${HOME}/.ollama}:/root/.ollama"

# Chat / embed models for ollama-pull (backend YAML is source of truth for the API).
x-soju-ai-chat-model: &soju-ai-chat-model gemma4:e4b
x-soju-ai-embed-model: &soju-ai-embed-model nomic-embed-text

# Speech: local = Soju backend TTS; browser = Web Speech API.
x-soju-tts-engine: &soju-tts-engine local
# Same-origin API via nginx (browser never talks to internal ports).
x-soju-api-base-url: &soju-api-base-url /api

# Minimal browser-visible env (models / tutor prompt come from GET /v1/soju/config/client).
x-soju-web-environment: &soju-web-environment
  DATA_DIR: /data
  PUBLIC_TTS_ENGINE: *soju-tts-engine
  PUBLIC_AI_ENABLED: "true"
  PUBLIC_AI_BASE_URL: *soju-api-base-url

name: soju

networks:
  soju:
    driver: bridge

services:
  web:
    build:
      context: .
      dockerfile: docker/Dockerfile.web
      target: dev
    # Compose command wins over image CMD so port changes apply without a rebuild.
    command: ["npm", "run", "dev", "--", "--host", "0.0.0.0", "--port", "14321"]
    expose:
      - "14321"
    volumes:
      - ./data:/data:rw
      - ./apps/web:/app:rw
      - web_node_modules:/app/node_modules
    environment:
      <<: *soju-web-environment
      # HMR websocket through published nginx :8080 (overridden in docker-compose.dev.yml).
      VITE_HMR_CLIENT_PORT: "8080"
    extra_hosts:
      - host.docker.internal:host-gateway
    networks:
      - soju

  backend:
    build:
      context: .
      dockerfile: docker/Dockerfile.backend
    command:
      [
        "soju",
        "backend",
        "--config",
        "/config/soju/backend.yaml",
        "--host",
        "0.0.0.0",
        "--port",
        "14322",
      ]
    volumes:
      - ./docker/soju:/config/soju:ro
      - ./data:/data:ro
    environment:
      DATA_DIR: /data
    extra_hosts:
      - host.docker.internal:host-gateway
    expose:
      - "14322"
    networks:
      - soju
    restart: unless-stopped

  docs:
    build:
      context: .
      dockerfile: docker/Dockerfile.docs
      target: prod
    expose:
      - "14323"
    networks:
      - soju
    restart: unless-stopped

  nginx:
    build:
      context: .
      dockerfile: docker/Dockerfile.nginx
    ports:
      - "8080:80"
    environment:
      SOJU_WEB_PORT: "8080"
    volumes:
      - ./docker/nginx/nginx.conf:/etc/nginx/conf.d/default.conf:ro
    depends_on:
      - web
      - backend
      - docs
    networks:
      - soju
    restart: unless-stopped

  ollama:
    image: ollama/ollama
    profiles: ["ollama"]
    environment:
      OLLAMA_ORIGINS: "*"
    ports:
      - "11434:11434"
    volumes:
      - *ollama-volume
    networks:
      - soju

  # Ensures the chat + embedding models exist in the shared Ollama cache (pull is incremental).
  ollama-pull:
    image: ollama/ollama
    profiles: ["ollama"]
    depends_on:
      ollama:
        condition: service_started
    environment:
      OLLAMA_HOST: http://ollama:11434
      OLLAMA_MODEL: *soju-ai-chat-model
      OLLAMA_EMBED_MODEL: *soju-ai-embed-model
    volumes:
      - *ollama-volume
    networks:
      - soju
    entrypoint: ["/bin/sh", "-c"]
    command:
      - |
        pull_with_retry() {
          model="$$1"
          for i in 1 2 3 4 5 6 7 8 9 10; do
            if ollama show "$$model" >/dev/null 2>&1; then
              echo "Model $$model already present"
              return 0
            fi
            ollama pull "$$model" && return 0
            sleep 3
          done
          echo "Failed to pull $$model" >&2
          return 1
        }
        status=0
        pull_with_retry "$$OLLAMA_MODEL" || status=1
        pull_with_retry "$$OLLAMA_EMBED_MODEL" || status=1
        exit "$$status"
    restart: "no"

volumes:
  web_node_modules: