Dockerfile

# Lexsi Data Science Agent — image for the Lexsi agent-deployment platform.
# The platform exposes ONE container via a public *.lexsi.ai URL, and that
# container MUST listen on port 7850 (host AND container). This image runs the
# Gradio app on 7850 (GRADIO_PORT below; the compose maps "7850:7850").
#
# Build for the platform host arch (linux/amd64) even on an Apple-silicon Mac,
# then push to Docker Hub (bplexsi):
#   docker buildx build --platform linux/amd64 \
#     -t bplexsi/lexsi-ds-agent:prod_v1 --push .

FROM python:3.10-slim

ENV PYTHONUNBUFFERED=1 \
    PIP_NO_CACHE_DIR=1 \
    # main() reads GRADIO_PORT; the app already binds server_name=0.0.0.0.
    GRADIO_PORT=7850 \
    GRADIO_ANALYTICS_ENABLED=False

WORKDIR /app

# curl is only needed for the compose healthcheck.
RUN apt-get update \
    && apt-get install -y --no-install-recommends curl \
    && rm -rf /var/lib/apt/lists/*

# --- Dependencies ---
# The pinned requirements.txt drags in GPU PyTorch + the full nvidia-*-cu13 /
# cuda-toolkit / triton stack (~10-15 GB) — useless without a GPU (the heavy ML
# runs on Lexsi pods). We keep sentence-transformers, pgserver, pymongo, etc.,
# but install CPU-only PyTorch from PyTorch's CPU wheel index so none of the
# CUDA packages come along. Two steps so torch is satisfied before
# sentence-transformers is resolved (otherwise pip would pull the CUDA wheel).
# CPU-only PyTorch. Use --extra-index-url (NOT --index-url) so torch's deps
# (typing-extensions, jinja2, fsspec, …) still resolve from PyPI; the explicit
# "+cpu" local version forces the CPU wheel (only on the pytorch index), so no
# CUDA/nvidia packages come along.
RUN pip install --extra-index-url https://download.pytorch.org/whl/cpu "torch==2.11.0+cpu"
# Everything else pinned to the tested lock (Python 3.10 variants) so the image
# behaves exactly like local. duckdb==1.5.2 is required — newer DuckDB's CSV
# reader fails to parse the headerless PKDD TSVs at build time.
RUN pip install \
    "gradio==5.50.0" \
    "lexsi-sdk==0.1.51" \
    "sentence-transformers==5.5.0" \
    "pgserver>=0.1.4" \
    "pymongo>=4.6" \
    "duckdb==1.5.2" \
    "pandas==2.3.3" \
    "numpy==2.2.6" \
    "pyyaml==6.0.3" \
    "networkx==3.4.2" \
    "rapidfuzz==3.14.5" \
    "pydantic==2.12.5" \
    "rich==15.0.0" \
    "typer==0.25.1"

# App code + the prebuilt artifacts/financial.duckdb (re-included in
# .dockerignore). The bundled PKDD source TSVs are Git-LFS pointers, so the DB
# can't be rebuilt at build time — we ship the already-built one instead.
COPY . .

# Sanity-check the bundled DuckDB opened correctly (fails the build early if the
# 35 MB file didn't get copied). At runtime `_ensure_pkdd_duckdb()` sees it
# present and skips any rebuild, so the Git-LFS TSVs are never parsed.
RUN python -c "import duckdb; duckdb.connect('artifacts/financial.duckdb', read_only=True).execute('select count(*) from fin_trans')"

EXPOSE 7850

# Runs app/gradio_app.py main() -> demo.launch(server_name=0.0.0.0, port=7850).
CMD ["python", "-m", "app.gradio_app"]