docker-compose.override.yml

# Enclave local-dev / self-host override (not part of upstream Onyx).
# Compose merges this file automatically on top of docker-compose.yml, so a
# plain `docker compose up -d` already includes everything below.
#
# It does two things:
#   1. Publishes api_server on :8080 so the Enclave app can reach the Onyx API
#      directly, without nginx + the Onyx web frontend (which boot-loops on the
#      published :latest image). nginx still serves the admin UI on :3001 once
#      the web frontend is healthy; :8080 is just a stable side door for the API.
#   2. Adds a local Ollama service so generation runs fully in-VPC — no
#      commercial LLM API. Onyx talks to it at http://ollama:11434 over the
#      compose network; 11434 is also published for `ollama` CLI / host debugging.
#
# ─────────────────────────────────────────────────────────────────────────────
# LLM RUNTIME — THREE WAYS TO RUN IT
# ─────────────────────────────────────────────────────────────────────────────
# In-stack Ollama is the DEFAULT because it ships as one `docker compose up`
# with no host-level install and behaves identically on Linux/macOS/Windows —
# the right trade-off for an open-source, self-hosted product. Pick a mode by
# how the host is resourced:
#
#   A. In-stack, CPU only  (default — works everywhere, slowest)
#        docker compose up -d
#      Good for any machine without a GPU, incl. Macs. CPU inference, so keep to
#      a small model (llama3.2:3b default; drop to :1b on tight memory).
#
#   B. In-stack, NVIDIA GPU  (Linux servers — the realistic VPC posture)
#        docker compose -f docker-compose.yml \
#                       -f docker-compose.override.yml \
#                       -f docker-compose.gpu.yml up -d
#      Adds GPU passthrough to the ollama service. Needs the NVIDIA Container
#      Toolkit on the host. Fast enough for llama3.1:8b and up. (Listing files by
#      hand disables Compose's auto-merge of this override, so it's listed too.)
#
#   C. Native Ollama on the host  (best on a Mac — Metal GPU, full host RAM)
#      Docker Desktop on macOS has NO GPU passthrough and caps container memory
#      (default ~8 GB), so in-stack is CPU-only and memory-bound there. Instead
#      run Ollama natively (`brew install ollama && ollama serve`), delete/ignore
#      the in-stack ollama service below, and re-seed Onyx to point at the host:
#        docker exec -e ENCLAVE_OLLAMA_API_BASE=http://host.docker.internal:11434 \
#          -e ENCLAVE_OLLAMA_DEFAULT_MODEL=llama3.1:8b \
#          onyx-api_server-1 python /tmp/enclave_seed_ollama.py
#
# ─────────────────────────────────────────────────────────────────────────────
# HARDWARE GUIDE (approx. RAM/VRAM the model needs, on top of the Onyx stack's
# ~6.5 GB; CPU inference is usable but slow, a GPU is ~10–50× faster)
# ─────────────────────────────────────────────────────────────────────────────
#   Model          Needs ~     Notes
#   llama3.2:1b    ~2 GB        Fits a default Docker Desktop; too weak for
#                              Onyx's agentic search prompt (emits garbage) —
#                              low-memory fallback only.
#   llama3.2:3b    ~4 GB        Default. Usable answers; raise Docker Desktop
#                              memory to ~12–16 GB before selecting it on a Mac.
#   llama3.1:8b    ~8 GB        Recommended for real use; comfortable on a GPU
#                              (mode B) or native on a 16 GB+ host (mode C).
# Switch model without editing code — the seed reads ENCLAVE_OLLAMA_MODELS /
# ENCLAVE_OLLAMA_DEFAULT_MODEL / ENCLAVE_OLLAMA_API_BASE (see enclave_seed_ollama.py).
services:
  api_server:
    ports:
      - "8080:8080"

  ollama:
    # Pinned by digest so the bundled runtime is reproducible (see
    # docker-compose.yml header for the version-control rationale).
    image: ollama/ollama@sha256:99262b6b2898e1d40907883e316f31e350e0ee6316ccae6127ac5a9feeacade2
    restart: unless-stopped
    ports:
      - "11434:11434"
    volumes:
      - ollama_data:/root/.ollama
    healthcheck:
      test: ["CMD", "ollama", "list"]
      interval: 10s
      timeout: 5s
      retries: 10

volumes:
  ollama_data: