pediatric-ai-scribe-v3/Dockerfile
Daniel 9d307fd442
Some checks failed
Forgejo Android APK / Root app tests (push) Successful in 50s
Forgejo Docker Build / Root app tests (push) Successful in 1m2s
Forgejo Android APK / Build signed APK (push) Successful in 2m6s
Forgejo Docker Build / Build Docker image (push) Successful in 37s
Forgejo Docker Build / Deploy to the host (push) Failing after 0s
feat: a vision model looks at the rendered deck and fixes the layout
The model that writes a deck never sees it. It cannot tell that slide four
overflowed, that a nine-item list would read better in two columns, or that two
labelled groups want to be a comparison — those are facts about the rendered
page, not about the text. So each generated deck is now rendered to PDF through
Gotenberg, rasterised to one image per slide with pdftoppm, and shown to a
vision model.

Off unless an administrator names a reviewer, in its own admin card because it
is the one setting that spends money on every generation without a user having
asked for anything. One pass, on generation only: a second pass costs as much as
the first and fixes far less, and refining is a text edit.

It returns a patch, not a deck. Asking for the corrected deck back put the reply
in proportion to the deck rather than to the number of problems, and a
fourteen-slide deck came back cut off mid-object at every output budget the
provider would honour — measured twice before changing shape.

The patch is better for a second reason. The reviewer names a slide and an
action — two columns, one column, split after bullet N, compare with these two
labels — and the server moves the text it already has. The words never pass
through the model, so a review cannot reword, drop or invent a single bullet.
That is a stronger guarantee than instructing it not to and checking afterwards.

The check runs anyway, because a bug in applyChanges would be as bad as a model
rewriting the words and worse for being trusted: body text must come out the
same multiset, figures the same set, and a heading may only be reused or
extended. A continuation heading is the reviewer's one piece of text and is
replaced when it does not continue anything.

Nothing here can fail a generation — no reviewer, an unreachable one, an
unparseable reply, a deck too long to look at, or a patch that applies to
nothing each return the deck that was written.

Verified end to end against a deck with a deliberately overloaded slide: three
slides rendered and sent, one change returned, ten bullets split into five and
five under "Stepwise Management … (continued)", text intact. Left switched off;
enable it under Admin.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
2026-09-11 21:38:18 +02:00

76 lines
4 KiB
Docker

# ─── OpenBao CLI, copied from upstream image (multi-arch automatic) ───
# Update the tag here to adopt a newer OpenBao. Binary is statically linked,
# safe to drop into the Node alpine image as-is.
# Pinned by digest, not by tag: a tag is a moving pointer, so two builds of the
# same commit could otherwise produce different images. These are manifest-list
# digests, so buildx still selects the right per-architecture variant.
FROM openbao/openbao:2.5.3@sha256:fdc6da21ca6963560c32336fd7feb9cf2d5e52668f1a1647205a4b41171f0806 AS bao-src
FROM node:24-alpine@sha256:e67514e5d0f6c46656005e1b693b2ec9d52e80b641307de684d4a015ba7a4eaf
WORKDIR /app
# ffmpeg: audio conversion for AWS Transcribe (WebM → PCM)
# curl: HTTP helper used by the OpenBao entrypoint and health/debug tooling
# jq: JSON parsing for the entrypoint's OpenBao secret-fetch step
# pandoc: Markdown → PPTX for Learning resources. It is large (~230MB), and it
# is here rather than in a sidecar because a sidecar would add a
# cross-stack network dependency to an export that must not fail for
# reasons outside this container. It also measures images, which
# pptxgenjs cannot: that library emits the target box verbatim with
# <a:stretch/>, so every image in every generated deck was distorted.
RUN apk add --no-cache ffmpeg curl jq pandoc-cli
# python-pptx builds the slide decks. pandoc still writes Word, where its output
# is good, but its pptx writer can only map markdown onto a handful of reference
# layouts: no per-slide layout, no positioning, no control over where an image
# lands or how large it is. That ceiling is the renderer's, not the model's — a
# better-written deck still came out as bullets on a template, and slides
# overflowed until autofit was injected into the emitted OOXML by hand.
#
# py3-lxml and py3-pillow come from apk rather than pip because both are C
# extensions and Alpine has no wheels for them; installing from source here
# would mean carrying a compiler in the runtime image. Adds ~58MB.
# poppler-utils supplies pdftoppm, which turns a rendered deck into one image
# per slide. That is the only way to let a vision model see what a deck actually
# looks like — Gotenberg converts to PDF and stops there.
RUN apk add --no-cache poppler-utils
RUN apk add --no-cache python3 py3-pip py3-lxml py3-pillow \
&& pip install --break-system-packages --no-cache-dir python-pptx==1.0.2 \
&& python3 -c 'import pptx'
# Pull the bao CLI out of the upstream image — matches host arch because
# buildx pulls the right manifest-list variant per build.
COPY --from=bao-src /bin/bao /usr/local/bin/bao
RUN /usr/local/bin/bao version
COPY package.json package-lock.json ./
# argon2 compiles native code via node-gyp — needs python3/make/g++ at build time
RUN apk add --no-cache --virtual .build-deps python3 make g++ \
&& npm ci --omit=dev \
&& apk del .build-deps
COPY . .
# One validated source revision for both runtime cache busting and OCI provenance.
# Direct development builds without an explicit revision remain visibly unversioned.
ARG GIT_REVISION=unknown
RUN node -e 'const r=process.argv[1]; if (r !== "unknown" && !require("./src/utils/buildId").isGitRevision(r)) throw new Error("GIT_REVISION must be a full lowercase Git SHA"); require("node:fs").writeFileSync("BUILD_ID", r + "\n");' -- "$GIT_REVISION"
LABEL org.opencontainers.image.revision=$GIT_REVISION
# Ensure the entrypoint is executable regardless of host file permissions
RUN chmod +x /app/docker-entrypoint.sh
RUN mkdir -p /app/data/logs
EXPOSE 3000
HEALTHCHECK --interval=30s --timeout=5s --start-period=20s \
CMD wget --no-verbose --tries=1 --spider http://localhost:3000/api/health || exit 1
# Entrypoint wrapper handles optional OpenBao secret fetch before exec'ing CMD.
# See docker-entrypoint.sh for the logic — it is a no-op if OPENBAO_ADDR is
# unset, so legacy .env-only deployments continue to work unchanged.
ENTRYPOINT ["/app/docker-entrypoint.sh"]
CMD ["node", "server.js"]