Dockerfile

# Build context is this directory:
#
#   docker build -t clair-algorithm .
#
# The image bundles torch (CPU wheel) + sentence-transformers and bakes the
# bge-m3 weights at build time, so a container boots without downloading ~2.2GB
# and needs no network access to Hugging Face at runtime. That makes the image
# large (~3GB) and the host RAM-heavy (the model wants several GB), which is a
# fair trade for a single long-lived process that should start fast offline.
FROM python:3.12-slim

WORKDIR /app

ENV PYTHONUNBUFFERED=1 \
    PIP_NO_CACHE_DIR=1 \
    HF_HOME=/opt/hf \
    HF_HUB_DISABLE_TELEMETRY=1 \
    EMBED_BACKEND=local

# CPU-only torch first, from PyTorch's CPU index — otherwise
# sentence-transformers pulls the default CUDA build and multi-GB of GPU
# libraries that a CPU host can't use.
RUN pip install torch --index-url https://download.pytorch.org/whl/cpu

# Then the rest (sentence-transformers finds torch already satisfied).
COPY requirements.txt ./
RUN pip install -r requirements.txt

# Bake the weights into the image's Hugging Face cache (HF_HOME).
RUN python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('BAAI/bge-m3')"

# With the weights cached, pin the runtime fully offline. Otherwise
# sentence-transformers HEAD-checks huggingface.co for model updates on every
# load; where egress to that host is slow or blocked, each check burns its full
# timeout per file and turns a fast local load into a multi-minute startup. Set
# after the bake, which does need the network.
ENV HF_HUB_OFFLINE=1 \
    TRANSFORMERS_OFFLINE=1

COPY . ./

CMD ["python", "main.py"]