compose.yml

# Reusable environment for Python/App services
x-app-env: &app-env
  UV_NO_SYNC: "1"
  HF_TOKEN: ${HF_TOKEN:-}
  HF_HUB_OFFLINE: "1"
  TRANSFORMERS_OFFLINE: "1"
  EMBED_MODEL: /model
  OLLAMA_HOST: ollama:11434
  OLLAMA_MODEL_ID: tinyllama
  CHAINLIT_FILES_DIR: /tmp/chainlit_files
  AGNAV_APP_NAME: ${AGNAV_APP_NAME:-BCGEU Navigator}
  AGNAV_APP_DESCRIPTION: ${AGNAV_APP_DESCRIPTION:-BCGEU Agreement Navigator}
  AGNAV_WELCOME_TITLE: ${AGNAV_WELCOME_TITLE:-BCGEU Navigator}
  AGNAV_DATA_DIR: ${AGNAV_DATA_DIR:-/data/sources}
  AGNAV_CACHE_DIR: ${AGNAV_CACHE_DIR:-/data/cache}

x-test-env: &test-env
  <<: *app-env

# Standard service boilerplate: Resource limits and Persistence
x-service-defaults: &service-defaults
  build:
    context: ./app
    dockerfile: Containerfile
    args:
      VERSION: ${AGNAV_VERSION:-dev}
    target: builder
  image: vexilon:dev
  mem_limit: 6g
  memswap_limit: 6g
  volumes: []

services:

  # ── Main Application
  # Development with Hot-Reload (Ollama)
  dev: &dev-service
    container_name: vexilon-dev
    <<: *service-defaults
    user: "${UID:-1000}:${GID:-1000}"
    depends_on:
      ollama-init:
        condition: service_completed_successfully
    volumes:
      - ./app:/app:z
      - /data/cache
      - /app/.files
    environment:
      <<: *app-env
      AGNAV_DATA_DIR: ${AGNAV_DATA_DIR:-/app/data}
      AGNAV_LLM_PROVIDER: ${AGNAV_LLM_PROVIDER:-ollama}
      AGNAV_DEV_MODE: "true"
      WATCHFILES_FORCE_POLLING: "true"
      PORT: "7860"
    mem_limit: 2g
    memswap_limit: 2g
    ports:
      - "7860:7860"
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:7860/"]
      interval: 10s
      timeout: 5s
      retries: 10
      start_period: 120s
    command: >
      sh -c "cd /app && uv run --no-sync chainlit run main.py --host 0.0.0.0 --port 7860 --headless --watch"

  # Staging / Cloud Parity test container running locally (Hugging Face)
  staging:
    container_name: vexilon-staging
    <<: *service-defaults
    user: "${UID:-1000}:${GID:-1000}"
    depends_on:
      ollama-init:
        condition: service_completed_successfully
    volumes:
      - ./app:/app:z
      - /data/cache
      - /app/.files
    environment:
      <<: *app-env
      AGNAV_DATA_DIR: ${AGNAV_DATA_DIR:-/app/data}
      AGNAV_LLM_PROVIDER: huggingface
      WATCHFILES_FORCE_POLLING: "true"
      PORT: "7861"
    mem_limit: 2g
    memswap_limit: 2g
    ports:
      - "7861:7861"
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:7861/"]
      interval: 10s
      timeout: 5s
      retries: 10
      start_period: 120s
    command: >
      sh -c "cd /app && uv run --no-sync chainlit run main.py --host 0.0.0.0 --port 7861 --headless --watch"

  # ── Tests
  # Pure logic tests with fully mocked external dependencies.
  # Fast, zero RAM footprint, runs on every commit.
  test-unit:
    container_name: vexilon-test-unit
    <<: *service-defaults
    environment: *test-env
    # Shared healthcheck not needed for one-off test runners
    command: >
      sh -c "VERIFY_ENABLED=true uv run --no-sync python -m pytest tests/test_*.py -v 
      --junitxml=reports/junit.xml 
      --cov=indexing --cov=main --cov-report=xml:reports/coverage.xml"

  # Validates integration between app code and the local model services.
  # Hits live Ollama instance; requires pre-warmed model weights.
  test-integration-model:
    container_name: vexilon-test-integration-model
    <<: *service-defaults
    depends_on:
      test-cache:
        condition: service_completed_successfully
      ollama-init-test:
        condition: service_completed_successfully
    environment: *test-env
    command: >
      sh -c "uv run --no-sync python -m pytest tests/integration/ -v"

  # Functional tests for the internal RAG streaming logic.
  # Verifies end-to-end Python flow from query to token with a mocked LLM.
  test-integration-app:
    container_name: vexilon-test-integration-app
    <<: *service-defaults
    depends_on:
      test-cache:
        condition: service_completed_successfully
    volumes: []
    environment:
      <<: *test-env
      PYTHONPATH: /app
      AGNAV_LLM_PROVIDER: mock
    command: >
      sh -c "uv run --no-sync python -m pytest tests/test_rag_stream.py -v"

  # CI-only app service without bind mount for e2e tests
  # Uses baked-in /app from image, avoiding bind-mount permission issues
  app-ci:
    container_name: vexilon-app-ci
    <<: *service-defaults
    user: "${UID:-1000}:${GID:-1000}"
    depends_on:
      ollama-init:
        condition: service_completed_successfully
    environment:
      <<: *app-env
      AGNAV_LLM_PROVIDER: ${AGNAV_LLM_PROVIDER:-ollama}
      PORT: "7860"
    mem_limit: 2g
    memswap_limit: 2g
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:7860/"]
      interval: 10s
      timeout: 5s
      retries: 10
      start_period: 120s
    command: >
      sh -c "cd /app && /venv-dev/bin/chainlit run main.py --host 0.0.0.0 --port 7860 --headless"

  # UI end-to-end tests using Playwright.
  # Verifies keyboard interactions, a11y labels, Knowledge Base drawer, and chat smokes.
  test-ui-e2e:
    container_name: vexilon-test-ui-e2e
    <<: *service-defaults
    depends_on:
      app-ci:
        condition: service_healthy
    environment:
      <<: *test-env
      APP_URL: http://app-ci:7860
      PLAYWRIGHT_BROWSERS_PATH: /ms-playwright
    command: >
      sh -c "uv run --no-sync playwright install chromium --with-deps && 
      uv run --no-sync python -m pytest tests/e2e/ -v"

  # Integrity gate for the pre-computed FAISS index.
  # Ensures the baked-in vector store matches the expected manifest.
  test-cache:
    container_name: vexilon-test-cache
    <<: *service-defaults
    environment: *app-env
    command: /venv/bin/python scripts/generate_cache_manifest.py validate

  # The final synchronization gate for the entire repository.
  # Only succeeds if all parallel unit and integration suites pass.
  test-everything:
    container_name: vexilon-test-everything
    <<: *service-defaults
    depends_on:
      test-unit:
        condition: service_completed_successfully
      test-integration-model:
        condition: service_completed_successfully
      test-integration-app:
        condition: service_completed_successfully
      test-ui-e2e:
        condition: service_completed_successfully
    command: >
      sh -c "echo 'ALL SYSTEMS VERIFIED'"

  # ── Local LLM (Ollama)
  # Core background daemon container. Runs 'ollama serve' indefinitely, loading
  # and serving the pulled models in RAM/VRAM on port 11434.
  ollama:
    container_name: vexilon-ollama
    image: ollama/ollama:0.40.5
    ports: ["${OLLAMA_PORT:-11434}:11434"]
    environment:
      OLLAMA_FLASH_ATTENTION: 1
      OLLAMA_HOST: 0.0.0.0
    volumes:
      - ${OLLAMA_DATA_PATH:-ollama_data}:/root/.ollama
    mem_limit: 4g
    memswap_limit: 4g
    healthcheck:
      test: ["CMD", "ollama", "list"]
      interval: 10s
      timeout: 5s
      retries: 5

  # One-time helper to pull the required model for local development.
  # Boots up, pulls model from registry, then exits immediately.
  # Fails loudly on network error so developer knows their dev env is broken.
  ollama-init: &ollama-init-service
    container_name: vexilon-ollama-init
    image: ollama/ollama:0.40.5
    depends_on:
      ollama:
        condition: service_healthy
    environment:
      <<: *app-env
    entrypoint: /bin/sh
    command: ["-c", "ollama pull $$OLLAMA_MODEL_ID"]

  # One-time helper to pull the model specifically for the testing environment.
  # Uses '*test-env' and has a safe '|| true' fallback: if registry is flaky/offline,
  # but model is already stored in host-restored CI cache, integration tests will still pass.
  ollama-init-test:
    <<: *ollama-init-service
    container_name: vexilon-ollama-init-test
    environment:
      <<: *test-env
    command: ["-c", "ollama pull $$OLLAMA_MODEL_ID || true"]

volumes:
  ollama_data: {}