Dockerfile

Rama main · réplica de lectura del repositorio original.

# Digest-pinned base image, refreshed by hand at every release (Debian security
# fixes): docker buildx imagetools inspect python:3.13-slim, then move the date.
# Base image checked on 2026-11-09
FROM python:3.13-slim@sha256:e48f59cdba69a3cbd9ed2cf0f348d90d45a8d350d98134cf1070bc106ee9ed90

RUN apt-get update && apt-get install -y --no-install-recommends \
    build-essential curl \
    && rm -rf /var/lib/apt/lists/*

WORKDIR /app

# Dependencies that contact a third party on their own (ONNX Runtime telemetry,
# Hugging Face Hub pings, LiteLLM's price table): off for every Python process
# in the image, the build-time model prefetch included. privaite sets the same
# defaults at import.
ENV ORT_DISABLE_TELEMETRY=1 \
    HF_HUB_DISABLE_TELEMETRY=1 \
    LITELLM_LOCAL_MODEL_COST_MAP=True

COPY pyproject.toml .
COPY privaite/ privaite/
RUN pip install --no-cache-dir .

RUN python -m spacy download en_core_web_lg && \
    python -m spacy download fr_core_news_md

# Run as a non-root user. Created BEFORE the model prefetch below so the Hugging
# Face cache is written into a directory that user owns; otherwise the container
# would re-download the model on every start.
RUN useradd --create-home --uid 10001 privaite
ENV HF_HOME=/home/privaite/.cache/huggingface
USER privaite

# Pre-download the default ONNX Privacy Filter model and tokenizer so the
# container starts fast and works offline from the first request.
# --build-arg ONNX_VARIANT=q4f16 bakes the other variant instead; the mounted
# config then needs the same pii.detectors.onnx.onnx_variant, or the first
# start downloads the default one.
ARG ONNX_VARIANT=""
RUN python -c "import os; from privaite.config.schema import OnnxDetectorConfig; from privaite.pii.detector_onnx import download_onnx_model; config = OnnxDetectorConfig(); download_onnx_model(repo_id=config.model_name, variant=os.environ.get('ONNX_VARIANT') or config.onnx_variant, revision=config.revision)" && \
    python -c "from privaite.config.schema import OnnxDetectorConfig; from transformers import AutoTokenizer; config = OnnxDetectorConfig(); AutoTokenizer.from_pretrained(config.model_name, revision=config.revision, trust_remote_code=config.trust_remote_code)"

# --build-arg MINILM=onnx (or onnx_int8) also bakes PrivAiTe's MiniLM model for
# preset: minilm, so that preset starts offline too. Off by default: the
# default preset does not use it, and it adds about 490 MB (120 MB for
# onnx_int8). The mounted config then needs the same pii.detectors.minilm.backend.
ARG MINILM=""
RUN if [ -n "$MINILM" ]; then \
      python -c "import os; from privaite.config.schema import MiniLMDetectorConfig; from privaite.pii.detector_minilm import BACKEND_FILES, download_minilm_files; config = MiniLMDetectorConfig(); download_minilm_files(config.model_name, ('config.json',) + BACKEND_FILES[os.environ['MINILM']], revision=config.revision)" && \
      python -c "from privaite.config.schema import MiniLMDetectorConfig; from transformers import AutoTokenizer; config = MiniLMDetectorConfig(); AutoTokenizer.from_pretrained(config.model_name, revision=config.revision)"; \
    fi

COPY config/ config/

EXPOSE 8400

# /ready, not /health: /health is a static dict that answers ok even when the
# PII engine is missing and every request 502s. /ready returns 503 until the
# engine is up. The start period covers the first-run model load.
HEALTHCHECK --interval=30s --timeout=5s --start-period=180s --retries=3 \
    CMD curl -f http://localhost:8400/ready || exit 1

# Config selection, in order: a mounted config/privaite.yaml wins; otherwise, if
# OPENAI_API_KEY is set, the drop-in OpenAI config is used (docker run -e
# OPENAI_API_KEY=... just works); otherwise the local Ollama example config, so
# the image still boots out of the box.
CMD ["sh", "-c", "if [ -f /app/config/privaite.yaml ]; then CFG=/app/config/privaite.yaml; elif [ -n \"$OPENAI_API_KEY\" ]; then CFG=/app/config/privaite.openai.yaml; else CFG=/app/config/privaite.example.yaml; fi; exec python -m privaite --config \"$CFG\""]