compose.openwebui.yaml

# Run Open WebUI on top of Seam's OpenAI-compatible gateway — one command, two containers.
#   docker model pull ai/gemma4:E2B  # once, so the model is available to Model Runner
#   docker compose -f compose.openwebui.yaml up --build -d
#   open http://localhost:3000
#
# Open WebUI talks to Seam as if it were OpenAI (/v1); Seam routes each request to the chosen
# backend. The model picker shows one entry per backend (<backend>:<model>).
services:
  app:
    build: .
    image: seam:latest
    environment:
      DEFAULT_BACKEND: docker
      # Reach the host's Docker Model Runner from inside the container via its gateway DNS.
      DMR_BASE_URL: http://model-runner.docker.internal/engines/v1
      # Default chat model. gemma4:E2B is the app default and returns native tool_calls (auto mode);
      # gemma3 does NOT (it emits an unparsed tool_code text block), so don't point the tool demo at it.
      DMR_MODEL: ai/gemma4:E2B
      # Real OpenAI key (so the `openai` backend / openai:gpt-4o-mini model works). Set it in the
      # gitignored .env (OPENAI_API_KEY=sk-...). Unset, the app boots with the openai backend
      # unconfigured (its models are omitted from /v1/models); supply a key later without a
      # restart:  PUT /admin/backends/openai/key  {"apiKey":"sk-..."}
      OPENAI_API_KEY: ${OPENAI_API_KEY:-}
    expose:
      - "8080"   # reachable by open-webui over the compose network; no host port needed

  open-webui:
    image: ghcr.io/open-webui/open-webui:v0.10.2
    ports:
      - "3000:8080"   # the UI: http://localhost:3000
    environment:
      # Point Open WebUI's "OpenAI API" connection at the Seam gateway.
      OPENAI_API_BASE_URL: http://app:8080/v1
      OPENAI_API_KEY: seam        # ignored by Seam, but Open WebUI requires a non-blank value
      ENABLE_OLLAMA_API: "false"     # OpenAI-compatible only; don't probe for Ollama
      WEBUI_AUTH: "false"            # skip the login/account step for a local demo
      # Optional: a HuggingFace token lifts the unauthenticated rate limit so the first-boot model
      # downloads are fast. Set the real value in a gitignored .env (HF_TOKEN=hf_...) — never here.
      HF_TOKEN: ${HF_TOKEN:-}
      # NOTE: on FIRST boot Open WebUI downloads its own models (a SentenceTransformers embedding
      # model for RAG, a faster-whisper model for audio) from HuggingFace — a few hundred MB that can
      # be slow on unauthenticated/rate-limited connections, and the container shows "unhealthy" until
      # it finishes. These are Open WebUI's own features (not Seam); they cache in the volume below,
      # so it only happens once. Chatting through the Seam gateway does not depend on them.
    depends_on:
      - app
    volumes:
      - open-webui:/app/backend/data

volumes:
  open-webui: