Dockerfile

# ─────────────────────────────────────────────────────────────────────────────
# Lexora API — one image, two targets.
#
#   Hugging Face Spaces  expects port 7860 and no card. Works as-is.
#   GCP Cloud Run        injects $PORT. The entrypoint honours it.
#
# The ONNX models are baked in at BUILD time, not fetched at boot. On a free tier
# a cold container that downloads ~200 MB before it can answer looks broken, and
# an outbound fetch on the startup path is one more thing that can fail in front
# of a user. The image is larger; the cold start is a model load, not a download.
#
# The portable index artefacts (chunks.jsonl + bm25.json) are copied in and the
# vector collection is materialised from them on first boot, so deployment never
# depends on two government web servers being up.
# ─────────────────────────────────────────────────────────────────────────────
FROM python:3.12-slim AS base

ENV PYTHONUNBUFFERED=1 \
    PYTHONDONTWRITEBYTECODE=1 \
    PIP_NO_CACHE_DIR=1 \
    HF_HUB_DISABLE_TELEMETRY=1 \
    LEXORA_VAR_DIR=/app/var

WORKDIR /app

# curl is used only by HEALTHCHECK below. tesseract is the offline OCR path for scanned
# uploads: Claude's vision model is better and is preferred when a key is present, but a
# demo without a key still has to read a photograph of a contract rather than refuse it.
RUN apt-get update \
 && apt-get install -y --no-install-recommends curl tesseract-ocr tesseract-ocr-eng \
 && rm -rf /var/lib/apt/lists/*

# Dependencies first so a code change does not invalidate the wheel layer.
COPY apps/api/pyproject.toml /app/apps/api/pyproject.toml
RUN pip install --upgrade pip && pip install /app/apps/api

COPY apps/api/app /app/apps/api/app
COPY corpus/sources.json corpus/manifest.json /app/corpus/
COPY var/index /app/var/index
# The recorded evaluation. Committed to the repo but, until now, never copied into the
# image — so /api/metrics found nothing and the "How well it works" page, the one an
# interviewer clicks first, rendered an empty state on every deploy.
COPY eval/results /app/eval/results

ENV PYTHONPATH=/app/apps/api

# Bake BOTH the models and the materialised vector store into the image.
#
# The models matter for cold-start latency. The vector store matters for whether the
# container starts at all: on a fresh filesystem the collection is empty, so the first
# boot would embed all 181 chunks before the health check could pass — tens of seconds
# on a warm laptop and minutes on a shared free-tier CPU, which reads to the platform as
# a failed deploy rather than a slow one.
#
# Doing it here moves that work to build time, where it is allowed to be slow. Startup
# then only opens an already-populated store.
#
# THREE layers, not one. A single process doing all three was OOM-killed by the Hugging
# Face build container (exit 137) with both ONNX sessions and all 181 vectors resident
# at once. The build container's ceiling is lower than the runtime's, so fitting in 1 GB
# at runtime — AUDIT.md §6.4 — does not imply the image can be built. Each stage exits
# before the next starts, so the kernel reclaims it and peak RSS is the largest stage
# rather than the sum. The artefacts are on disk, so splitting the process does not
# split the result. Each stage prints its peak RSS; see scripts/bake.py.
COPY scripts/bake.py /app/scripts/bake.py

RUN python scripts/bake.py embedding
RUN python scripts/bake.py reranker
RUN python scripts/bake.py vectorstore

# Run as a non-root user. Spaces and Cloud Run both allow it, and there is no
# reason for an inference service to be root.
RUN useradd --create-home --uid 1000 lexora \
 && chown -R lexora:lexora /app
USER lexora

EXPOSE 7860
HEALTHCHECK --interval=30s --timeout=5s --start-period=90s --retries=3 \
  CMD curl -fsS "http://127.0.0.1:${PORT:-7860}/api/health" || exit 1

# Cloud Run sets $PORT; Hugging Face Spaces expects 7860.
CMD ["sh", "-c", "exec uvicorn app.main:app --host 0.0.0.0 --port ${PORT:-7860} --workers 1"]