# =============================================================================
# Custom LiteLLM image — fastText language detection (privacy-plus).
# -----------------------------------------------------------------------------
# The privacy-plus language-detect step runs inside litellm/privacy_scoring.py,
# which lives in the LiteLLM pod — NOT in the Presidio one. fastText is a COMPILED
# dependency + a model file, so it cannot ride in the ConfigMap that carries the
# .py hooks — it is baked here. The hook .py themselves STILL come from the
# ConfigMap (mounted at /app/litellm), so you can iterate on them without
# rebuilding this image.
#
# The fastText LID model lid.176.ftz is licensed CC BY-SA 3.0.
#
# Prepare the build context first (the model is NOT committed to git):
#   curl -LO https://dl.fbaipublicfiles.com/fasttext/supervised-models/lid.176.ftz
#
# Prebuilt: quay.io/asalvati/litellm-privacy-plus:0.1
# =============================================================================
FROM ghcr.io/berriai/litellm-non_root:v1.91.1

USER 0

# fastText runtime. NOTE: this base image (Wolfi + uv venv, Python 3.13) ships NO pip,
# so bootstrap it with ensurepip first. We use `fasttext-predict` (predict-only, MIT):
# it publishes prebuilt cp313 manylinux wheels, so NO C++ toolchain is needed — unlike
# `fasttext` / `fasttext-wheel`, which lack a cp313 wheel and try to compile from source.
RUN python -m ensurepip && python -m pip install --no-cache-dir fasttext-predict

# fastText language-id model: 176 languages, ~917 KB. Fetched at BUILD time (build
# host has egress); the running pod stays no-egress. Path matches
# policy/privacy-plus.yaml -> ner.lang_detect.model_path.
COPY lid.176.ftz /opt/models/lid.176.ftz

# Restore the base image's non-root user (65534 / nobody).
USER 65534