Dockerfile

# Inkference — HF Docker Space for the FULL corpus (all 6 books, ~923 pages).
#
# Differences from the Book-1 Dockerfile:
#   * Bakes the PREBUILT DB + FAISS index (app/deploy/seed_data) — no boot re-seed,
#     so cold start is instant even for 923 pages.
#   * Page scans are NOT baked (1.3 GB). They're served by redirecting to a public
#     HF dataset via the INKFERENCE_IMAGES_BASE_URL Space variable.
#   * Live upload (Kraken + TrOCR + confidence + Groq correction) still works.
#
# HF builds the repo-ROOT Dockerfile, so the deploy script copies this to ./Dockerfile.
FROM python:3.11-slim

RUN apt-get update && apt-get install -y --no-install-recommends \
        libgl1 libglib2.0-0 && rm -rf /var/lib/apt/lists/*
RUN useradd -m -u 1000 user
WORKDIR /app

# Full pipeline deps (Kraken + TrOCR + RAG + API) so live upload works.
COPY app/deploy/requirements-space.txt .
RUN pip install --no-cache-dir -r requirements-space.txt

COPY app/pyproject.toml ./
COPY app/src/ ./src/
COPY app/frontend/ ./frontend/
RUN pip install --no-cache-dir -e . --no-deps && \
    mkdir -p /data && chown -R user:user /app /data

ENV INKFERENCE_DATA_ROOT=/data \
    HF_HOME=/home/user/.cache/huggingface \
    PORT=7860 \
    INKFERENCE_LOG_LEVEL=INFO \
    # recognition (live upload)
    TROCR_MODEL_ID=microsoft/trocr-base-handwritten \
    HTR_MAX_LONG_EDGE=1600 \
    HTR_NUM_BEAMS=1 \
    # post-correction via Groq (GROQ_API_KEY secret)
    CORRECTION_ENABLED=true \
    CORRECTION_BACKEND=api \
    CORRECTION_API_BASE=https://api.groq.com/openai/v1 \
    CORRECTION_API_MODEL=qwen/qwen3-32b \
    # Ask-the-Archive: Groq -> Gemini fallback -> extractive
    LLM_PROVIDER=groq \
    LLM_MODEL=openai/gpt-oss-120b \
    LLM_FALLBACK=gemini:gemini-2.5-flash-lite \
    RAG_USE_CORRECTED=true \
    # dataset holding the prebuilt corpus (seed_data/) AND the scans (book*/forster*/)
    SEED_DATASET=sajitkun125/inkference-book-images
# REQUIRED Space secrets: GROQ_API_KEY, GEMINI_API_KEY
# REQUIRED Space variable (page scans): INKFERENCE_IMAGES_BASE_URL=
#   https://huggingface.co/datasets/<user>/inkference-book-images/resolve/main
# (uploaded pages save absolute paths under /data and are served directly.)

USER user

# Pre-download embedding (RAG) + base recognition (upload) models for a fast first request.
RUN python -c "from sentence_transformers import SentenceTransformer as S; S('sentence-transformers/all-MiniLM-L6-v2')"
RUN python -c "from transformers import TrOCRProcessor, VisionEncoderDecoderModel as M; TrOCRProcessor.from_pretrained('microsoft/trocr-base-handwritten'); M.from_pretrained('microsoft/trocr-base-handwritten')"

# Fetch the prebuilt DB + FAISS index from the dataset. snapshot_download RESOLVES LFS
# to real bytes (unlike the Space's Docker build, which ships LFS as pointers). Only the
# seed_data/ folder is pulled — the 1.3 GB scans stay remote and stream via the CDN.
RUN python -c "import os; from huggingface_hub import snapshot_download; \
snapshot_download(os.environ['SEED_DATASET'], repo_type='dataset', \
allow_patterns=['seed_data/*'], local_dir='/app/hub')"

EXPOSE 7860

# Copy the prebuilt corpus into the (ephemeral) data root on boot, then serve. No re-seed.
CMD ["sh", "-c", "cp -rn /app/hub/seed_data/. /data/ 2>/dev/null || true; uvicorn inkference.api.main:app --host 0.0.0.0 --port ${PORT:-7860}"]