# Quick start: local model (llama.cpp) and the Constructor gateway.
#   cp docker/env.example docker/.env   (set both keys)
#   docker compose -f docker/compose.yml up -d
# The first start downloads the model (about 2.7 GB). CPU only, 8 GB RAM are enough.
services:
  llama:
    # llama.cpp build 11783 (the project was measured with 11724), pinned by digest
    image: ghcr.io/ggml-org/llama.cpp:server@sha256:c7b03e45cff9c4e984acd39032a6f3a82c811075b72bf176e329f84b8598c135
    command:
      - -hf
      - unsloth/Qwen3.5-4B-GGUF:Q4_K_M
      - --host
      - 0.0.0.0
      - --port
      - "8080"
      - --ctx-size
      - "8192"
      - --parallel
      - "1"
      - --jinja
      # Text only: do not load the image projector that -hf would fetch along
      - --no-mmproj
      # Thinking off: otherwise the model thinks freely first and the JSON schema only applies afterwards
      - --reasoning
      - "off"
    environment:
      LLAMA_API_KEY: ${LLAMA_API_KEY:?set LLAMA_API_KEY in docker/.env}
      LLAMA_CACHE: /models
    volumes:
      - models:/models
    restart: unless-stopped

  gateway:
    build:
      context: ..
      dockerfile: docker/gateway.Dockerfile
    environment:
      GATEWAY_API_KEY: ${GATEWAY_API_KEY:?set GATEWAY_API_KEY in docker/.env}
      LLAMA_API_KEY: ${LLAMA_API_KEY}
      LLAMA_URL: http://llama:8080
      # Seconds per model call. The first call processes the long system prompt; on a slow CPU that takes minutes
      LLAMA_TIMEOUT: "900"
    ports:
      # Reachable from your Luanti server. Every request needs the API key.
      - "8000:8000"
    volumes:
      - data:/data
    depends_on:
      - llama
    restart: unless-stopped

volumes:
  models:
  data: