# localchat + a local Gemma 4 E2B model served by llama.cpp.
#
#   docker compose up -d        # first start downloads the model (~3 GB)
#   open http://localhost:3000
#
# Swap the model with LLM_HF_REPO, e.g. LLM_HF_REPO=unsloth/gemma-4-E4B-it-GGUF:Q4_K_M
services:
  llm:
    image: ghcr.io/ggml-org/llama.cpp:server
    command:
      - -hf
      - ${LLM_HF_REPO:-unsloth/gemma-4-E2B-it-GGUF:Q4_K_M}
      - --no-mmproj      # text-only chat; skip the vision projector download
      - --ctx-size
      - "8192"
      - --port
      - "8080"
    environment:
      LLAMA_CACHE: /models
    volumes:
      - models:/models
    healthcheck:
      test: ["CMD", "curl", "-fs", "http://localhost:8080/health"]
      interval: 10s
      timeout: 5s
      start_period: 15m   # first boot downloads the model
    restart: unless-stopped

  app:
    build: .
    image: localchat:latest
    ports:
      - "3000:3000"
    environment:
      LLM_BASE_URL: http://llm:8080/v1
      LOCALCHAT_TITLE: ${LOCALCHAT_TITLE:-localchat}
      SYSTEM_PROMPT: ${SYSTEM_PROMPT:-}
    depends_on:
      llm:
        condition: service_healthy
    restart: unless-stopped

volumes:
  models: