Dockerfile

# Playwright's official image ships Chromium + all system deps prematched to the
# playwright pip version below — avoids hand-rolling apt/xvfb setup for headless Chrome.
FROM mcr.microsoft.com/playwright/python:v1.44.0-jammy

WORKDIR /app

# Fetch + unpack the Keyword Surfer extension at build time. Non-fatal if it fails —
# modules/keywords.py falls back to Autocomplete/pytrends/LLM estimates at runtime.
COPY scripts/fetch_extension.py ./scripts/fetch_extension.py
RUN pip install --no-cache-dir httpx && \
    python scripts/fetch_extension.py --out /app/extension || true

# CPU-only torch, installed BEFORE requirements.txt so sentence-transformers'
# transitive "torch" dependency is already satisfied by this build instead of
# pulling PyPI's default CUDA-enabled build (~2GB of unused nvidia-cu* wheels
# for a GPU this Space's default CPU hardware doesn't have — the RAG embedder
# here is a small model doing a handful of queries per request, CPU is fine).
RUN pip install --no-cache-dir torch --index-url https://download.pytorch.org/whl/cpu

COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt

COPY app.py .
COPY modules/ ./modules/
COPY data/ ./data/

ENV GRADIO_SERVER_NAME=0.0.0.0
ENV GRADIO_SERVER_PORT=7860
# Without this, stdout is block-buffered in a non-TTY container: print()
# output (our own startup/progress logs) sits unflushed while stderr-based
# output (tqdm progress bars) appears immediately, making logs look stuck.
ENV PYTHONUNBUFFERED=1
EXPOSE 7860

# rag_index/ is NOT copied at build time — it's 2.3GB (over the free-tier 1GB
# private-repo LFS limit baked into an image would mean re-transferring it on
# every build). Instead modules/rag.py downloads it at container startup from
# a separate private HF Dataset repo, via RAG_DATASET_ID + RAG_DATASET_TOKEN.
#
# Required as Space secrets in Settings > Repository secrets, NOT here and NOT
# via a committed .env file:
#   RAG_DATASET_ID          e.g. vivekchakraverty/digital-marketer-rag-index
#   RAG_DATASET_TOKEN        read access to that dataset repo
#   GOOGLE_ADS_DEVELOPER_TOKEN, GOOGLE_ADS_CLIENT_ID, GOOGLE_ADS_CLIENT_SECRET,
#   GOOGLE_ADS_REFRESH_TOKEN, GOOGLE_ADS_LOGIN_CUSTOMER_ID   (optional tier-1
#   keyword data; modules/keywords.py falls back to tier 2 if unset)
# HF Spaces injects Repository secrets as environment variables at container
# runtime automatically — no Dockerfile/ENV changes needed to receive them.

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