bf4f895f2c
The ceiling was 100 MB per file with ten files allowed at once, and every file is held whole in memory to be parsed — so the old limit let a single request ask for a gigabyte of heap. A source article that size is not a thing anyone uploads here. Now 10 MB, defined once and used by both the multer limit and the post-upload check. The filter accepted `allowed mime OR allowed extension`, so naming a file .pdf was enough on its own, whatever it declared — and the extension is chosen by whoever uploads. Both are required now. Neither of those sees any bytes: multer filters on the headers, before the file has arrived. verifySources() runs once the buffer exists and refuses a file whose contents are not what its type claims, using the same helper as documents, S3 uploads and assistant attachments. It runs before extraction, because an extractor handed a malformed file is where the damage would happen. The CMS screen said 100 MB and listed four of the ten accepted formats; it now says what the server actually does. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Dv6sqaY6Vq3ChZHMem3cnU
4 KiB
Learning Hub
A CMS + content-delivery module for clinical education material inside the app. Supports articles, clinical pearls, quizzes, and Marp-rendered presentations with PPTX export. Quiz questions are stored alongside article content and can optionally be generated by AI from uploaded source material.
Content types
| Type | Description |
|---|---|
article |
Rich HTML body with an optional attached quiz |
pearl |
Short clinical snippet (no quiz, no heavy media) |
quiz |
Standalone quiz (no article body) |
presentation |
Marp markdown rendered as slides; PPTX export supported |
User-facing features
- Browse by category.
- Three search modes:
- Keyword — Postgres full-text.
- Semantic — pgvector cosine similarity on the embedding column.
- Hybrid — weighted merge of both result sets.
- Articles render with sanitized HTML (DOMPurify, loaded via SRI-pinned cdnjs).
- Quizzes: multiple-choice, multi-select, true/false. Score computed on submit, per-question explanations revealed after.
- Presentation viewer: modal with keyboard / swipe navigation.
- Progress:
learning_progressstores per-attempt score + total.
CMS (moderator / admin)
- Tiptap rich-text editor for article body.
- Draft / published toggle.
- Category assignment.
- Quiz builder: add/remove questions, add/remove options, mark correct, enter explanation.
- Marp editor for presentations with live preview.
AI content generation
POST /api/admin/learning/generate takes one of:
| Input | Notes |
|---|---|
topic |
Plain-text description of the topic |
| Uploaded files | PDF / DOCX / PPTX / ODT / EPUB / TXT / MD / HTML / CSV / JSON, ≤ 10 MB each, max 10 files. The declared type must be in the allowlist and match the extension, and the bytes are sniffed before anything parses them. |
| WebDAV path | Pulled from the user's connected Nextcloud instance |
Parameters: model (from the provider whitelist), slideCount for
presentations, wordCount for articles.
File uploads pass the src/utils/fileType.js magic-byte check so a
mismatched extension is rejected before it reaches the parser.
Marp → PPTX export
POST /api/admin/learning/generate-pptx writes the markdown to a temp
directory and runs pandoc against
assets/learning/slides-reference.pptx. The reference deck carries the fonts,
palette and slide layouts, so restyling the export means editing that file in
PowerPoint — not changing code.
Images are handled before pandoc sees the markdown. Each
/api/generated-images/{id} link is resolved through the ownership check and
written beside the deck under a name this route chooses; any link that does not
resolve to one of those is dropped. Pandoc resolves image links against the
filesystem, so passing an arbitrary local path through would embed that file
into the deck.
This is the Learning Hub's own path and is separate from My Resources, which
renders decks with python-pptx from a typed deck rather than from markdown —
see my-resources.md.
Semantic search
| Store | pgvector on learning_content.embedding VECTOR(768) |
| Index | IVFFLAT, cosine distance |
| Primary model | vertex/text-embedding-005 (768 dims), served through LiteLLM |
| Fallback model | OpenAI text-embedding-3-small (truncated to 768 to match the column) |
Embeddings are generated on content publish + on every edit. If the embedding provider is unreachable, the content still saves — keyword search remains available.
Tables
| Table | Purpose |
|---|---|
learning_categories |
Top-level groupings |
learning_content |
Articles / pearls / quizzes / presentations. Body + embedding vector. |
learning_questions |
Quiz question prompts (FK to content) |
learning_options |
Answer options (FK to question) |
learning_progress |
Per-user attempt history |
Retrieval sizing
How many corpus excerpts the Clinical Assistant, the Learning Hub and My Resources each receive, and the reranker cap that overrides all three: retrieval-tuning.md.