Dockerfile

# SIMOSphere AI Copilot Mini — self-contained ExApp image.
# Bakes the embedding model (e5-large, CPU) AND the prebuilt Nextcloud-docs index into the image,
# so the app runs fully offline: the only outbound call at runtime is to the user's chosen LLM.
FROM python:3.12-slim-bookworm

ENV PYTHONUNBUFFERED=1 \
    PIP_NO_CACHE_DIR=1 \
    HF_HOME=/app/models \
    COPILOT_MINI_DATA=/app/data \
    APP_HOST=0.0.0.0 \
    APP_PORT=9034

WORKDIR /app

# CPU-only torch first, so the heavy CUDA wheels are never pulled in by sentence-transformers.
RUN pip install torch --index-url https://download.pytorch.org/whl/cpu

# Python deps from the project metadata.
COPY pyproject.toml README.md ./
COPY src/ ./src/
RUN pip install "nc_py_api>=0.30" . \
 && python -c "import copilot_mini; print(copilot_mini.__version__)" \
 && python -c "import copilot_mini, pathlib; p=pathlib.Path(copilot_mini.__file__).parent/'static'/'js'/'main.js'; assert p.is_file(), 'FIX #2 regression: static/ not bundled into the installed package: %s' % p; print('static bundled OK:', p)"

# Bake the embedding model into the image (network only during build). Only the safetensors
# weights are pulled — the repo also ships PyTorch-.bin, TensorFlow-.h5 and ONNX copies (~2.2 GB
# each) that sentence-transformers does not need; skipping them cuts the image by ~6 GB.
RUN python -c "from huggingface_hub import snapshot_download; snapshot_download('intfloat/multilingual-e5-large', ignore_patterns=['*.bin','*.h5','*.onnx','*.msgpack','*.ot','tf_model*','flax_model*','rust_model*','onnx/*','openvino/*'])"

# Prebuilt, self-contained knowledge index (CC BY 3.0 Nextcloud docs).
COPY data/ ./data/

# Run offline from here on — model + index are baked in.
ENV HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1

EXPOSE 9034
ENTRYPOINT ["python", "-m", "copilot_mini.app"]