compose.yaml

# Compose orchestration for Adrian.
#
# Run from the repo root after `adrian-setup bootstrap` has populated
# .env and seeded data/adrian.db. Bring the stack up (backend +
# dashboard + bundled classifier; requires NVIDIA Container Toolkit):
#
#   docker compose --profile llm up -d
#
# One-shot bootstrap / reset / set-model use the setup profile:
#
#   docker compose --profile setup run --rm setup bootstrap --gguf gemma-4-e2b.gguf
#   docker compose --profile setup run --rm setup reset-password
#   docker compose --profile setup run --rm setup set-model --gguf gemma-4-e4b.gguf

services:
  backend:
    build:
      context: .
      dockerfile: deploy/Dockerfile.backend
    image: adrian-backend:local
    # Build-only: never pull from a registry. Stops a future
    # squatter who registers `adrian-backend:local` on Docker Hub
    # from getting silently pulled and run.
    pull_policy: build
    container_name: adrian-backend
    restart: unless-stopped
    # .env is optional so `docker compose --profile setup run setup
    # bootstrap` can run on a fresh clone before any .env exists.
    # The bootstrap subcommand writes .env; subsequent `docker
    # compose up` calls then find it.
    env_file:
      - path: .env
        required: false
    environment:
      ADRIAN_DB_PATH: /data/adrian.db
      # Pin the in-container listen to 8080. ADRIAN_BACKEND_PORT in
      # .env / env_file only controls the host-side mapping below; if
      # we let the Go process read it too it would try to bind that
      # host port inside its own netns and the published mapping would
      # forward to a port nothing was listening on.
      ADRIAN_BACKEND_PORT: "8080"
    ports:
      - "${ADRIAN_BACKEND_PORT:-8080}:8080"
    volumes:
      - ./data:/data
      - ./models:/models:ro
    # Distroless image has no shell / wget / curl, so the binary
    # self-probes via `/adrian healthcheck` (HTTP-GET /readyz, exit
    # 0/1). start_period covers the cold path: first DB open + the
    # classifier upstream's first response (Llama.cpp model load can
    # take 30-60 s on a fresh container).
    healthcheck:
      test: ["CMD", "/adrian", "healthcheck"]
      interval: 10s
      timeout: 5s
      retries: 5
      start_period: 60s
    # Hold the backend off until the classifier is ready - eliminates
    # the early-window where a SDK could connect, fail to classify, and
    # (since the engine fails closed) halt every tool call.
    depends_on:
      llm:
        condition: service_healthy
        required: false

  setup:
    profiles: ["setup"]
    build:
      context: .
      dockerfile: deploy/Dockerfile.setup
    image: adrian-setup:local
    pull_policy: build
    volumes:
      - .:/workspace

  frontend:
    build:
      context: .
      dockerfile: deploy/Dockerfile.frontend
    image: adrian-frontend:local
    pull_policy: build
    container_name: adrian-frontend
    restart: unless-stopped
    depends_on:
      - backend
    ports:
      - "${ADRIAN_DASHBOARD_PORT:-3000}:3000"

  # Local classifier. Enable with the "llm" profile:
  #   docker compose --profile llm up -d
  #
  # Hosts a GGUF model behind a Chat Completions endpoint at
  # http://adrian-llm:8081/v1/chat/completions (the ADRIAN_LLM_URL
  # value `bootstrap --gguf <name>` writes).
  #
  # NVIDIA GPU is required; the upstream
  # ghcr.io/ggml-org/llama.cpp:server-cuda image bundles CUDA + the
  # llama-server binary.
  llm:
    profiles: ["llm"]
    image: ghcr.io/ggml-org/llama.cpp:server-cuda
    container_name: adrian-llm
    restart: unless-stopped
    env_file:
      - path: .env
        required: false
    volumes:
      - ./models:/models:ro
    ports:
      - "127.0.0.1:8081:8081"
    # ADRIAN_LLM_MODEL_PATH is set by `setup bootstrap --gguf <name>` to
    # the in-container path (e.g. /models/gemma-4-e2b.gguf). n-gpu-layers
    # 99 offloads everything to the GPU; llama.cpp clamps to the model's
    # actual layer count, so this is a safe upper bound.
    command: >-
      -m ${ADRIAN_LLM_MODEL_PATH:-/models/model.gguf}
      --host 0.0.0.0
      --port 8081
      --ctx-size ${ADRIAN_LLM_CTX_SIZE:-8192}
      --n-gpu-layers 99
      --jinja
    # llama-server exposes /health (returns 200 once the model is
    # loaded and ready to accept requests) and /v1/models. The
    # upstream image is Ubuntu-based and ships curl, so a shell-form
    # check works. start_period covers the load: a 5 GB GGUF + GPU
    # offload routinely takes 30-60 s on a cold container.
    healthcheck:
      test: ["CMD-SHELL", "curl -fsS http://localhost:8081/health || exit 1"]
      interval: 10s
      timeout: 5s
      retries: 12
      start_period: 90s
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]