# llama-cag-n8n v2 — fully containerized CAG stack.
# All ports bind to 127.0.0.1. Configure via .env (python llamacag.py setup).
# GPU: add -f docker-compose.gpu.yml (or: python llamacag.py start --gpu).

services:
  llama-server:
    image: ${LLAMA_IMAGE:-ghcr.io/ggml-org/llama.cpp:server}
    container_name: llamacag-llama
    restart: unless-stopped
    # The "local-llama" profile keeps this service in the default all-in-Docker
    # setup. Native mode (Apple Silicon / Metal, where Docker has no GPU
    # passthrough) runs llama-server on the host instead: set COMPOSE_PROFILES=
    # empty in .env and point LLAMA_SERVER_URL at host.docker.internal.
    # See `python llamacag.py start --native-llama` and docs/HARDWARE.md.
    profiles: ["local-llama"]
    # -hf downloads the GGUF from Hugging Face into the llama_models volume on
    # first boot (~6.5 GB for the default model), then runs offline.
    # q8_0 KV cache quantization halves context memory at negligible quality
    # cost (flash attention is auto-enabled on supported backends).
    # Cache-persistence guards (keep all four, in all three compose files):
    # --swa-full is REQUIRED for saved slot states to restore correctly on
    # sliding-window-attention models (all Gemma) and needs a server built on/
    # after 2026-04-24 — it sizes KV by the full context on every layer (no SWA
    # discount), which is the RAM price of trustworthy restores. --no-mmproj,
    # --cache-ram 0 and --ctx-checkpoints 0 keep the multimodal projector and
    # the newer host-RAM cache/checkpoint paths (where the open SWA cache bugs
    # live) away from slot persistence. Evidence: docs/ROADMAP.md F16.
    command: >
      -hf ${LLAMA_MODEL:-google/gemma-4-12B-it-qat-q4_0-gguf}
      --ctx-size ${LLAMA_CTX_SIZE:-65536}
      --parallel ${CAG_SLOTS:-1}
      --cache-type-k ${LLAMA_CACHE_TYPE_KV:-q8_0}
      --cache-type-v ${LLAMA_CACHE_TYPE_KV:-q8_0}
      --slot-save-path /caches
      --swa-full
      --no-mmproj
      --cache-ram 0
      --ctx-checkpoints 0
      --host 0.0.0.0
      --port 8080
      --threads ${LLAMA_THREADS:--1}
      ${LLAMA_EXTRA_ARGS:-}
    environment:
      - LLAMA_CACHE=/models
    volumes:
      - llama_models:/models
      - kv_caches:/caches
    ports:
      - "127.0.0.1:${LLAMA_PORT:-8080}:8080"

  cag-api:
    build: ./api
    container_name: llamacag-api
    restart: unless-stopped
    environment:
      # Defaults to the in-Docker service; override in .env with
      # LLAMA_SERVER_URL=http://host.docker.internal:8080 for native mode.
      - LLAMA_SERVER_URL=${LLAMA_SERVER_URL:-http://llama-server:8080}
      - LLAMA_CTX_SIZE=${LLAMA_CTX_SIZE:-65536}
      - CAG_SLOTS=${CAG_SLOTS:-1}
      # Stack-wide answer defaults (per-request values still win). Chatbot
      # deployments typically pin DEFAULT_TEMPERATURE=0.0 for deterministic,
      # extractive answers — see "Running it as a dedicated chatbot" in README.
      - DEFAULT_TEMPERATURE=${DEFAULT_TEMPERATURE:-0.2}
      - DEFAULT_MAX_ANSWER_TOKENS=${DEFAULT_MAX_ANSWER_TOKENS:-1024}
      # Paraphrase tolerance for POST /verify's quote grounding (higher = stricter).
      - QUOTE_MATCH_THRESHOLD=${QUOTE_MATCH_THRESHOLD:-0.9}
      # Cloud per-1k input-token price for GET /stats' savings estimate (0.0 hides it).
      - CLOUD_PRICE_PER_1K_INPUT=${CLOUD_PRICE_PER_1K_INPUT:-0.0}
      # POST /documents/{id}/calibrate: battery cap + non-strict scoring pass line.
      - CALIBRATE_MAX_ITEMS=${CALIBRATE_MAX_ITEMS:-100}
      - CALIBRATE_MATCH_THRESHOLD=${CALIBRATE_MATCH_THRESHOLD:-0.85}
      # Serve the zero-install web UI at /ui (loopback-only; unauthenticated stack).
      - WEBUI_ENABLED=${WEBUI_ENABLED:-true}
      - MAX_UPLOAD_MB=${MAX_UPLOAD_MB:-50}
      # llama call timeouts (seconds) — raise QUERY_TIMEOUT_S on slow CPUs if
      # first-after-restart queries on big documents hit the 502 wall.
      - QUERY_TIMEOUT_S=${QUERY_TIMEOUT_S:-600}
      - WARM_TIMEOUT_S=${WARM_TIMEOUT_S:-3600}
      - CACHE_DIR=/caches
      - DB_HOST=db
      - DB_PORT=5432
      - DB_NAME=${DB_NAME:-llamacag}
      - DB_USER=${DB_USER:-llamacag}
      - DB_PASSWORD=${DB_PASSWORD:?set DB_PASSWORD in .env (python llamacag.py setup)}
    volumes:
      - kv_caches:/caches
    ports:
      - "127.0.0.1:${CAG_API_PORT:-8000}:8000"
    depends_on:
      db:
        condition: service_healthy
      # required: false so cag-api still starts in native mode, when the
      # llama-server service is disabled (its "local-llama" profile is off) and
      # inference runs on the host. cag-api reaches it via LLAMA_SERVER_URL.
      llama-server:
        condition: service_started
        required: false

  n8n:
    image: docker.n8n.io/n8nio/n8n:${N8N_VERSION:-latest}
    container_name: llamacag-n8n
    restart: unless-stopped
    environment:
      - N8N_PORT=5678
      - N8N_PROTOCOL=http
      - N8N_HOST=${N8N_HOST:-localhost}
      - N8N_SECURE_COOKIE=false
      - N8N_DIAGNOSTICS_ENABLED=false
      - N8N_PERSONALIZATION_ENABLED=false
      - N8N_RUNNERS_ENABLED=true
      - N8N_ENCRYPTION_KEY=${N8N_ENCRYPTION_KEY:?set N8N_ENCRYPTION_KEY in .env (python llamacag.py setup)}
      - N8N_USER_MANAGEMENT_JWT_SECRET=${N8N_USER_MANAGEMENT_JWT_SECRET:?set N8N_USER_MANAGEMENT_JWT_SECRET in .env (python llamacag.py setup)}
      - GENERIC_TIMEZONE=${GENERIC_TIMEZONE:-UTC}
      - DB_TYPE=postgresdb
      - DB_POSTGRESDB_HOST=db
      - DB_POSTGRESDB_PORT=5432
      - DB_POSTGRESDB_DATABASE=n8n
      - DB_POSTGRESDB_USER=${DB_USER:-llamacag}
      - DB_POSTGRESDB_PASSWORD=${DB_PASSWORD}
    volumes:
      - n8n_data:/home/node/.n8n
      # Watched by the document-ingestion workflow (Local File Trigger).
      - ${DOCUMENTS_FOLDER:-./documents}:/data/documents
    ports:
      - "127.0.0.1:${N8N_PORT:-5678}:5678"
    depends_on:
      db:
        condition: service_healthy

  db:
    image: postgres:17-alpine
    container_name: llamacag-db
    restart: unless-stopped
    environment:
      - POSTGRES_USER=${DB_USER:-llamacag}
      - POSTGRES_PASSWORD=${DB_PASSWORD}
      - POSTGRES_DB=${DB_NAME:-llamacag}
    volumes:
      - postgres_data:/var/lib/postgresql/data
      # 01: creates the separate `n8n` database; 02: llamacag app schema.
      - ./database/init-n8n-db.sh:/docker-entrypoint-initdb.d/01-init-n8n-db.sh:ro
      - ./database/schema.sql:/docker-entrypoint-initdb.d/02-schema.sql:ro
    ports:
      - "127.0.0.1:${DB_PORT:-5432}:5432"
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U ${DB_USER:-llamacag} -d ${DB_NAME:-llamacag}"]
      interval: 5s
      timeout: 5s
      retries: 10

volumes:
  n8n_data:
  postgres_data:
  llama_models:
  kv_caches: