docker-compose.cuda.yml

services:
  code-index-api:
    image: dvcdsys/code-index:cu128
    # No container_name on purpose. A fixed name is global to the docker
    # daemon, so a second cix on the same host (a Portainer stack, another
    # clone, a re-run from a different directory) dies with "the container
    # name /code-index is already in use". Compose's own
    # <project>-<service>-1 naming keeps them side by side; ask for the real
    # name with `docker compose ps` when a command needs it.
    restart: unless-stopped
    ports:
      - "${PORT:-21847}:21847"
    environment:
      - CIX_API_KEY=${CIX_API_KEY}
      # Defense in depth — the image already defaults to 21847 but
      # pinning it here keeps the host:container port mapping honest
      # if a third-party fork or custom build sets a different default.
      - CIX_PORT=${CIX_PORT:-21847}
      - CIX_EMBEDDING_MODEL=${CIX_EMBEDDING_MODEL:-awhiteside/CodeRankEmbed-Q8_0-GGUF}
      # Legacy chromem-go store: read once on startup for the one-time import
      # into the SQLite vector store, then left untouched as the rollback path.
      - CIX_CHROMA_PERSIST_DIR=/data/chroma
      # Optional. Vector store: one SQLite database per embedding namespace.
      # Defaults to a sibling of CIX_CHROMA_PERSIST_DIR (/data/vectors), so it
      # already lands on the volume below — set it only to move the vectors to
      # a different mount.
      # - CIX_VECTORS_DIR=/data/vectors
      # Optional. PRAGMA mmap_size for the vector store, in bytes; 0 (default)
      # is off. Roughly 40% lower search latency in exchange for resident
      # memory — mapped database pages count in RSS, per connection. There is
      # room for it under the 10G limit below on a GPU host.
      # - CIX_VECTOR_MMAP_SIZE=2147483648
      - CIX_SQLITE_PATH=/data/sqlite/projects.db
      - CIX_MAX_FILE_SIZE=${CIX_MAX_FILE_SIZE:-524288}
      - CIX_EXCLUDED_DIRS=${CIX_EXCLUDED_DIRS:-node_modules,.git,.venv,__pycache__,dist,build,.next,.cache,.DS_Store}
      - CIX_N_GPU_LAYERS=99
      # GGUF cache lives on the named volume below — survives `docker compose
      # down` (without -v) and is owned by the image's 1001:1001 user, so the
      # cix-server process can always write to it regardless of host
      # bind-mount permissions.
      - CIX_GGUF_CACHE_DIR=/data/models
      - CIX_LLAMA_BIN_DIR=/app
      - CIX_LLAMA_STARTUP_TIMEOUT=120
      - CIX_EMBEDDINGS_ENABLED=${CIX_EMBEDDINGS_ENABLED:-true}
      # ── First-boot admin seed (required when the DB has no users yet) ──
      # cix-server refuses to start when the users table is empty AND these
      # are unset. Set BOTH in your .env, log in once, change the password
      # immediately (the user is flagged must_change_password=true).
      - CIX_BOOTSTRAP_ADMIN_EMAIL=${CIX_BOOTSTRAP_ADMIN_EMAIL:-}
      - CIX_BOOTSTRAP_ADMIN_PASSWORD=${CIX_BOOTSTRAP_ADMIN_PASSWORD:-}
      # ── PR-E runtime tunables (all DB-overridable from /dashboard/server) ──
      # 0 = auto. Threads → runtime.NumCPU()/2; batch → match n_ctx.
      - CIX_LLAMA_THREADS=${CIX_LLAMA_THREADS:-0}
      - CIX_LLAMA_BATCH=${CIX_LLAMA_BATCH:-0}
      # Embedding queue parallelism. Default 5 (was 1) — pipelines host-side
      # prep with device inference. Drop to 1 if you observe contention.
      - CIX_MAX_EMBEDDING_CONCURRENCY=${CIX_MAX_EMBEDDING_CONCURRENCY:-5}
      - CIX_EMBEDDING_QUEUE_TIMEOUT=${CIX_EMBEDDING_QUEUE_TIMEOUT:-300}
      # Optional: skip the first-boot HF download by pointing at a GGUF
      # file the operator already has on disk. cix copies it into the
      # cix-models named volume once (atomic .partial → rename) and never
      # touches the source again. Subsequent boots find the file in cache
      # and ignore the env. See volumes block below for an example bind.
      - CIX_BOOTSTRAP_GGUF_PATH=${CIX_BOOTSTRAP_GGUF_PATH:-}
      # ── Pluggable embedding providers (added in migration 12) ──
      # The active provider is selected from the dashboard
      # (/dashboard/server → Embedding provider). On first boot with an
      # empty runtime_settings row, cix-server seeds the ollama provider
      # using the CIX_EMBEDDING_MODEL + CIX_LLAMA_* vars above — so the
      # default deployment is unchanged.
      #
      # To use a remote provider (OpenAI-compatible or Voyage AI) you
      # MUST export the API-key env var below. The dashboard reads it on
      # every embed call; cix-server NEVER persists API keys in the DB,
      # only the env-var NAME a provider should look up. Switching
      # providers triggers a full reindex per project on the next clone
      # job (existing model-change pipeline).
      #
      # OpenAI-compatible (api.openai.com, vLLM, TEI, LocalAI, …):
      # - CIX_OPENAI_API_KEY=${CIX_OPENAI_API_KEY:-}
      #
      # Voyage AI:
      # - CIX_VOYAGE_API_KEY=${CIX_VOYAGE_API_KEY:-}
      - NVIDIA_VISIBLE_DEVICES=all
    volumes:
      # Operator-managed bind for sqlite + chroma so backups and inspection
      # are one `cd` away on the host. On Linux the directory must be owned
      # by 1001:1001 (`sudo chown -R 1001:1001 ~/.cix/data`) or the server
      # cannot create its database; `user: "0:0"` is the less safe fallback.
      # install-server.sh does the chown for you. See doc/TEAM_DEPLOYMENT.md.
      - ${HOME}/.cix/data:/data
      # Docker-managed named volume layered ON TOP of /data/models. This
      # isolates the GGUF cache from host-side bind permission issues and
      # guarantees the model is downloaded exactly once across container
      # recreates (`docker compose up --force-recreate`, image bumps, etc.).
      - cix-models:/data/models
      # Optional bootstrap: bind a host-side .gguf read-only into
      # /bootstrap/model.gguf and set CIX_BOOTSTRAP_GGUF_PATH=/bootstrap/model.gguf
      # in your .env. cix imports it into the cix-models cache on first boot,
      # then ignores both the env and the bind. After verifying the cache is
      # seeded, the bind can be removed entirely.
      # - /srv/hf-cache/coderankembed-q8_0.gguf:/bootstrap/model.gguf:ro
    deploy:
      resources:
        limits:
          memory: 10G
        reservations:
          memory: 2G
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    # /health only answers once the whole boot is done — llama model load
    # (minutes from cold) and, on the first start after an upgrade, the
    # one-time chromem -> SQLite vector-store import. Failures inside
    # start_period do not count, so this window is simply "how long before a
    # slow boot may be called broken". Nothing acts on `unhealthy` here
    # (restart: unless-stopped reacts to exits, not health), so an over-long
    # window costs nothing, while a short one paints the container red mid-boot
    # and invites a restart that throws the boot away.
    healthcheck:
      test: ["CMD", "/cix-server", "-healthcheck"]
      interval: 30s
      timeout: 10s
      start_period: 600s
      retries: 3

volumes:
  cix-models:
    # GGUF model cache. Persisted by Docker; only `docker compose down -v`
    # (or explicit `docker volume rm <project>_cix-models`) wipes it.