# TalkToDoc production image.
# Built to be interchangeable across hosts: Hugging Face Spaces, Render
# (Docker deploy), Fly.io, Cloud Run, or any other platform that runs a
# standard Docker image. Nothing here is tied to one specific host.
# Configuration comes entirely from environment variables set by
# whichever platform runs it (PORT, DATABASE_URL, OPENAI_API_KEY,
# SECRET_KEY, ENVIRONMENT).

FROM python:3.11-slim

# Whisper needs ffmpeg to read audio files, this installs it at the
# system level since it can't come from pip. build-essential is needed
# because some packages in requirements.txt compile C extensions during
# install, and the slim base image has no compiler by default.
RUN apt-get update && apt-get install -y --no-install-recommends \
    build-essential \
    ffmpeg \
    && rm -rf /var/lib/apt/lists/*

# Run as a non-root user, standard practice and specifically expected by
# Hugging Face Spaces' Docker SDK.
RUN useradd -m -u 1000 user
USER user
ENV PATH="/home/user/.local/bin:$PATH"

WORKDIR /home/user/app

COPY --chown=user requirements.txt .
RUN pip install --no-cache-dir --upgrade -r requirements.txt

# YarnGPT downloads its WavTokenizer checkpoint itself on first import,
# using a bare requests.get() call with no error checking. If that
# download hiccups at runtime (inside the running container, on the
# app's first request), it silently saves a corrupted or partial file
# and the app fails later with a confusing, unrelated error.
#
# This downloads the exact same two files ahead of time, at build time,
# using huggingface_hub's properly tested download function instead. If
# this fails, the build fails clearly, right here, instead of the app
# failing mysteriously after it's already live. YarnGPT's own downloader
# checks whether the file already exists before downloading, so it will
# find these already in place and skip straight past its own fragile
# download step.
RUN python - <<'PY'
from huggingface_hub import hf_hub_download
import os

target = "/home/user/.yarngpt/models"
os.makedirs(target, exist_ok=True)

hf_hub_download(
    repo_id="novateur/WavTokenizer-large-speech-75token",
    filename="wavtokenizer_large_speech_320_24k.ckpt",
    revision="c5512c1bd34afef082035923e4bae245f3da9e5f",
    local_dir=target,
)

hf_hub_download(
    repo_id="novateur/WavTokenizer-medium-speech-75token",
    filename="wavtokenizer_mediumdata_frame75_3s_nq1_code4096_dim512_kmeans200_attn.yaml",
    local_dir=target,
)

print("WavTokenizer files downloaded.")
PY

COPY --chown=user . .

ENV ENVIRONMENT=production

# 7860 is Hugging Face Spaces' default port. Other platforms (like
# Render) set $PORT themselves, and this falls back to that automatically.
EXPOSE 7860

CMD gunicorn app:app --bind 0.0.0.0:${PORT:-7860} --timeout 120 --workers 1