| # App image. CPU-only torch; BGE and the cross-encoder are downloaded at | |
| # build time so the container never fetches models at startup. | |
| FROM python:3.12-slim | |
| ENV PYTHONUNBUFFERED=1 PIP_NO_CACHE_DIR=1 | |
| # torch from the CPU index first: from PyPI, sentence-transformers would pull the | |
| # CUDA build (several GB) on x86_64. | |
| RUN pip install --index-url https://download.pytorch.org/whl/cpu torch==2.14.1 | |
| COPY requirements-app.txt /tmp/ | |
| RUN pip install -r /tmp/requirements-app.txt | |
| # HF Spaces runs the container as uid 1000; build as that user so the model | |
| # cache and the app files are its own. | |
| RUN useradd -m -u 1000 user | |
| USER user | |
| WORKDIR /home/user/app | |
| RUN python -c "from sentence_transformers import CrossEncoder, SentenceTransformer; \ | |
| SentenceTransformer('BAAI/bge-small-en-v1.5'); \ | |
| CrossEncoder('cross-encoder/ms-marco-MiniLM-L6-v2')" | |
| ENV HF_HUB_OFFLINE=1 PYTHONPATH=/home/user/app/src | |
| COPY --chown=user src/multimodal_rag/ src/multimodal_rag/ | |
| COPY --chown=user data/chunks.jsonl data/ | |
| COPY --chown=user streamlit_app.py . | |
| # No file watcher: nothing to hot-reload in a container, and walking every | |
| # loaded module made transformers log ~2,400 lines of torchvision import errors. | |
| EXPOSE 7860 | |
| CMD ["streamlit", "run", "streamlit_app.py", "--server.port=7860", \ | |
| "--server.address=0.0.0.0", "--server.headless=true", \ | |
| "--server.fileWatcherType=none", "--client.toolbarMode=viewer"] | |