Dockerfile

# Negoptim AI backend — built for Hugging Face Spaces (Docker Space, free CPU tier)
#
# Build strategy: everything heavy happens at BUILD time so the free tier's
# ephemeral disk doesn't matter at runtime —
#   1. CPU-only torch (several GB smaller than the default CUDA build)
#   2. the e5 embedding model is pre-downloaded into the image
#   3. the knowledge base is ingested into ChromaDB inside the image
# Runtime then cold-starts in seconds with the index already in place.
#
# Secrets (set in the Space settings, never in this file):
#   GROQ_API_KEY   (required)
#   CORS_ORIGINS   e.g. https://your-app.vercel.app,http://localhost:3000
#   SMTP_USER / SMTP_PASSWORD          (optional, real email delivery)
#   CEREBRAS_API_KEY / GEMINI_API_KEY  (optional, extra LLM fallback)

FROM python:3.11-slim

# HF Spaces runs containers as UID 1000
RUN useradd -m -u 1000 user
USER user
ENV PATH="/home/user/.local/bin:$PATH" \
    HF_HOME=/home/user/.cache/huggingface \
    PYTHONUNBUFFERED=1

WORKDIR /app

# CPU-only torch first, so sentence-transformers doesn't pull the CUDA build
COPY --chown=user backend/requirements.txt .
RUN pip install --no-cache-dir --user torch --index-url https://download.pytorch.org/whl/cpu && \
    pip install --no-cache-dir --user -r requirements.txt

# Pre-download the embedding model into the image layer cache
RUN python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('intfloat/multilingual-e5-small')"

COPY --chown=user backend/ /app/backend/
COPY --chown=user knowledge/ /app/knowledge/
COPY --chown=user scripts/ /app/scripts/

WORKDIR /app/backend
ENV CHROMA_PERSIST_DIRECTORY=/app/backend/chroma_db \
    KNOWLEDGE_BASE_PATH=/app/knowledge

# Build the vector index into the image (no persistent disk needed at runtime)
RUN python ../scripts/ingest.py

# HF Spaces expects the app on port 7860
EXPOSE 7860
CMD ["uvicorn", "app.main:app", "--host", "0.0.0.0", "--port", "7860"]