docker-compose.quickstart.yml

# EdgeQuake — One-Command Quickstart
#
# Pulls all three services from prebuilt GHCR images.
# No Rust toolchain, no Node.js, no local build step required.
#
# Usage:
#   docker compose -f docker-compose.quickstart.yml up -d
#
# Or via Makefile:
#   make stack
#
# Access:
#   Web UI  → http://localhost:3000
#   API     → http://localhost:8080
#   Swagger → http://localhost:8080/swagger-ui
#   Health  → http://localhost:8080/health
#
# LLM provider (default: Ollama on the host at port 11434)
# To use OpenAI:
#   EDGEQUAKE_LLM_PROVIDER=openai OPENAI_API_KEY=sk-... \
#     docker compose -f docker-compose.quickstart.yml up -d
#
# Pin to a specific release:
#   EDGEQUAKE_VERSION=0.10.6 docker compose -f docker-compose.quickstart.yml up -d

services:
  # ── PostgreSQL with pgvector + Apache AGE ──────────────────────────────────
  postgres:
    image: ghcr.io/raphaelmansuy/edgequake-postgres:${EDGEQUAKE_VERSION:-latest}
    container_name: edgequake-postgres
    restart: unless-stopped
    shm_size: '256m'
    environment:
      POSTGRES_USER:     edgequake
      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:-edgequake_secret}
      POSTGRES_DB:       edgequake
    volumes:
      - edgequake-pg-data:/var/lib/postgresql/data
    networks:
      - edgequake
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U edgequake -d edgequake"]
      interval: 10s
      timeout: 5s
      retries: 5
      start_period: 10s

  # ── EdgeQuake API ──────────────────────────────────────────────────────────
  api:
    image: ghcr.io/raphaelmansuy/edgequake:${EDGEQUAKE_VERSION:-latest}
    container_name: edgequake-api
    restart: unless-stopped
    ports:
      - "${EDGEQUAKE_PORT:-8080}:8080"
    environment:
      EDGEQUAKE_HOST: 0.0.0.0
      EDGEQUAKE_PORT: 8080
      DATABASE_URL: postgres://edgequake:${POSTGRES_PASSWORD:-edgequake_secret}@postgres:5432/edgequake
      # ── LLM Provider ────────────────────────────────────────────────────────
      # Default: Ollama running on the host machine.
      # Override: EDGEQUAKE_LLM_PROVIDER=openai OPENAI_API_KEY=sk-...
      EDGEQUAKE_LLM_PROVIDER:       ${EDGEQUAKE_LLM_PROVIDER:-ollama}
      # Optional model override — when unset the API picks a sensible default
      # for the resolved provider (e.g. gpt-5-mini for OpenAI, gemma4 for Ollama).
      EDGEQUAKE_LLM_MODEL:          ${EDGEQUAKE_LLM_MODEL:-}
      # ── Embedding Provider ───────────────────────────────────────────────────
      # Default: same provider as LLM.  Set to "openai" when OPENAI_API_KEY is
      # present and you want OpenAI embeddings regardless of the LLM provider.
      EDGEQUAKE_EMBEDDING_PROVIDER: ${EDGEQUAKE_EMBEDDING_PROVIDER:-}
      # Optional model override — when unset the API picks a sensible default
      # (e.g. text-embedding-3-small for OpenAI, embeddinggemma for Ollama).
      EDGEQUAKE_EMBEDDING_MODEL:    ${EDGEQUAKE_EMBEDDING_MODEL:-}
      # ── Vision LLM Provider (PDF → Markdown) ────────────────────────────────
      # Used exclusively for rendering PDF pages into Markdown (multimodal LLM).
      # Default: same provider as EDGEQUAKE_LLM_PROVIDER (set by quickstart.sh).
      # Override: EDGEQUAKE_VISION_PROVIDER=openai to use OpenAI even if Ollama
      #           is the main entity-extraction LLM.
      # WHY nested substitution: Docker Compose ${VAR:-} maps an unset host var to
      # the empty string "" inside the container. std::env::var then returns Ok("")
      # which looks like a valid value and bypasses fallback chains in the server code.
      # Using ${VISION_VAR:-${LLM_VAR:-default}} ensures the container always has a
      # non-empty value derived from the corresponding LLM setting when not explicit.
      EDGEQUAKE_VISION_PROVIDER:    ${EDGEQUAKE_VISION_PROVIDER:-${EDGEQUAKE_LLM_PROVIDER:-ollama}}
      # Vision model: when unset falls back to the main LLM model so Ollama users
      # automatically use their configured model (e.g. gemma4:e4b) for vision too.
      EDGEQUAKE_VISION_MODEL:       ${EDGEQUAKE_VISION_MODEL:-${EDGEQUAKE_LLM_MODEL:-}}
      OPENAI_API_KEY:               ${OPENAI_API_KEY:-}
      # WHY: OPENAI_BASE_URL must NOT be set to an empty string.
      # When unset on the host ${OPENAI_BASE_URL:-} evaluates to "", which the
      # OpenAI provider uses as the base URL, breaking every request with
      # "builder error".  The API server strips empty env vars at startup, but
      # setting a safe default here provides defence-in-depth.
      # Override only when using a custom OpenAI-compatible endpoint
      # (e.g. OPENAI_BASE_URL=http://localhost:11434/v1 for Ollama in compat mode).
      OPENAI_BASE_URL:              ${OPENAI_BASE_URL:-}
      ANTHROPIC_API_KEY:            ${ANTHROPIC_API_KEY:-}
      GEMINI_API_KEY:               ${GEMINI_API_KEY:-}
      MISTRAL_API_KEY:              ${MISTRAL_API_KEY:-}
      # WHY host.docker.internal (NOT localhost):
      #   Docker containers have their own network namespace. 'localhost' inside
      #   a container refers to the container itself, NOT to the host where Ollama
      #   runs. 'host.docker.internal' is the canonical name for the host machine:
      #     macOS/Windows: provided natively by Docker Desktop
      #     Linux:         mapped via extra_hosts below (host-gateway)
      #
      # IMPORTANT: If you set OLLAMA_HOST manually, never use 'localhost' or
      #   '127.x.x.x' — use 'host.docker.internal' or a real IP/hostname.
      #   quickstart.sh auto-translates loopback addresses when launching.
      OLLAMA_HOST:                  ${OLLAMA_HOST:-http://host.docker.internal:11434}
      # WHY: Pass through OLLAMA_CONTEXT_LENGTH so users can tune the model
      # token window without rebuilding. Default 32768; set to 131072 for large docs.
      OLLAMA_CONTEXT_LENGTH:        ${OLLAMA_CONTEXT_LENGTH:-32768}
      OLLAMA_MODEL:                 ${OLLAMA_MODEL:-}
      OLLAMA_EMBEDDING_MODEL:       ${OLLAMA_EMBEDDING_MODEL:-}
      PDFIUM_AUTO_CACHE_DIR:        /tmp/edgequake-pdfium-cache
      RUST_LOG:                     ${RUST_LOG:-info}
    extra_hosts:
      # Maps host.docker.internal → host machine IP on Linux.
      # On macOS/Windows, Docker Desktop provides this DNS name natively.
      - "host.docker.internal:host-gateway"
    depends_on:
      postgres:
        condition: service_healthy
    networks:
      - edgequake
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8080/health"]
      interval: 20s
      timeout: 10s
      retries: 5
      start_period: 15s

  # ── Next.js Web UI ─────────────────────────────────────────────────────────
  frontend:
    image: ghcr.io/raphaelmansuy/edgequake-frontend:${EDGEQUAKE_VERSION:-latest}
    container_name: edgequake-frontend
    restart: unless-stopped
    ports:
      - "${FRONTEND_PORT:-3000}:3000"
    environment:
      NODE_ENV: production
      # WHY EDGEQUAKE_API_URL (not NEXT_PUBLIC_API_URL):
      # NEXT_PUBLIC_* variables are baked into the JS bundle at image build time
      # and cannot be changed at container startup. EDGEQUAKE_API_URL is read at
      # request time by the Next.js server component (layout.tsx → getRuntimeConfig)
      # and injected into window.__EDGEQUAKE_RUNTIME_CONFIG__ so the browser picks
      # it up correctly — no rebuild required.
      #
      # Default: http://localhost:${EDGEQUAKE_PORT:-8080} — the URL the user's
      # browser uses to reach the API (same host, host-mapped port).
      #
      # Override for remote/non-localhost access:
      #   EDGEQUAKE_API_URL=http://192.168.1.10:8080 docker compose ... up -d
      EDGEQUAKE_API_URL:  ${EDGEQUAKE_API_URL:-http://localhost:${EDGEQUAKE_PORT:-8080}}
    depends_on:
      api:
        condition: service_healthy
    networks:
      - edgequake
    healthcheck:
      test: ["CMD", "wget", "--no-verbose", "--tries=1", "--spider", "http://localhost:3000"]
      interval: 20s
      timeout: 10s
      retries: 3
      start_period: 20s

volumes:
  edgequake-pg-data:
    driver: local

networks:
  edgequake:
    driver: bridge