If Campo knew what Campo knows.
An AI-readable markdown corpus built from the public parts of FAU Campo, fau.de's Studiengang pages, and FAU's Prüfungsordnungen (PDFs converted to markdown). The name riffs on the old Siemens saying "if Siemens knew what Siemens knows" — Campo already exposes a lot, but its JSF UI makes it hard to actually find anything. We turn that public surface into a flat list of merged-per-program markdown files an LLM agent can navigate.
Status: SoSe 2026 corpus complete. One merged file per Campo study program (FAU.de Studiengang content + every PO-version + every course inlined), plus the full PO regulation texts as separate large files. Weekly + monthly crons keep it fresh.
Each weekly run cuts a GitHub Release tagged snapshot-YYYY-Www with ifcampoknew-corpus-{period-slug}.zip as the asset — a single self-contained markdown corpus ready to drop into a RAG system. Start with README.md inside the zip; it explains the layout and stable-id conventions to the agent.
A self-contained markdown corpus + the Python tooling that builds it. Three public sources, one merged corpus:
data/
├── README.md # RAG usage guide (also at the zip root)
├── {period-slug}/ # Campo course catalogue per semester
│ ├── INDEX.md # programs grouped by section
│ └── {program-slug-id}.md # ★ ONE merged file per program:
│ # • FAU.de Studiengang inline (Steckbrief, sections)
│ # • every PO-version with permalinks + dated PDFs
│ # • every course (Eckdaten + Termine + instructors)
│ # • Lehramts-Prüfungsordnungen (when applicable)
├── studiengang/ # FAU.de Studiengang pages (raw, also inlined above)
│ └── {slug}.md
└── pruefungsordnungen/ # full PO regulation texts (10–30 k tok each)
└── {faculty}/{program}/{po-slug}.md
Slugs are Campo/FAU-faithful German with ASCII-folded umlauts (ä→ae, ß→ss); every Campo basename ends in -<segmentId> so links survive renames. Every file carries campo_permalink / source_url / pdf_source in its YAML front-matter so a RAG can cite the original source.
Original plan was a Web Components SPA on GitHub Pages. After analysis we realised: the LLM agent already does what an SPA's UI does. Drop the UI, ship the data in the form an agent eats best — markdown — and let agents (or github.com's native renderer) be the front end. Web Components, FlexSearch, onboarding quiz, telemetry: all gone. The corpus is the product.
| Path | What's in it |
|---|---|
data/{period-slug}/ |
Campo per-semester corpus (catalogue tree + course detail). The primary deliverable. |
data/studiengang/ |
FAU.de Studiengang info — one markdown per program (222). |
data/pruefungsordnungen/ |
FAU.de regulations — landing pages + each PO converted from PDF. |
scraper/ |
Python pipeline. See scraper/README.md. |
docs/campo-public-surface.md |
Inventory of every Campo page/flow reachable without logging in, plus the working POST recipe and deep-link patterns we use. |
docs/requirements.md |
v2-locked requirements (markdown-corpus pivot, decisions log, open items). |
devlog.md |
Chronological log of every prompt → actions → findings → time spent. |
tmp/ |
(gitignored) JSON snapshots and downloaded PDFs, regenerated each run. |
# install
python3 -m venv scraper/.venv
scraper/.venv/bin/pip install -r scraper/requirements.txt
# Campo (per-semester catalogue + course detail + merged-program render)
scraper/.venv/bin/python scraper/scrape.py --period 589 --out tmp/589.json --max-depth 4 -v
scraper/.venv/bin/python scraper/fetch_courses.py --in tmp/589.json --out tmp/589-courses.json -v
scraper/.venv/bin/python scraper/render_markdown.py --in tmp/589.json --courses tmp/589-courses.json --out data
# FAU.de (programs + Prüfungsordnungen — single-script pipeline)
scraper/.venv/bin/python scraper/fau_corpus.py --out data --tmp tmp/fau-pdfs -v
# Re-render once you have FAU files on disk so they get inlined:
scraper/.venv/bin/python scraper/render_markdown.py --in tmp/589.json --courses tmp/589-courses.json --out data
# Build a corpus zip locally
( cd data && zip -r -q ../ifcampoknew-corpus.zip . )Browse the result under data/.
| Phase | Status | What it covers |
|---|---|---|
| Public-surface analysis | ✅ | What Campo exposes anonymously; HTTP/JSF mechanics; deep-link patterns. |
| Requirements (v1 → v2 pivot) | ✅ | v1 was a Web Components UI; v2 is the markdown corpus. |
| Catalogue scrape (Campo) | ✅ | depth-4 BFS: 1 895 nodes for SoSe 2026, with checkpoint/resume. |
| Course-content attachment | ✅ | Fetch every Course's Termine + Eckdaten + Lehrende; 683/683 for SoSe 2026. |
| FAU.de Studiengang corpus | ✅ | 222 programs scraped; content inlined into the matching Campo program file. |
| FAU.de Prüfungsordnungen corpus | ✅ | 36 landing pages + ~2 800 PDFs converted to markdown via PyMuPDF4LLM. |
| Cross-link Campo ↔ FAU.de | ✅ | Each program file inlines its FAU.de Studiengang page and references the matched dated PO-PDFs. |
| Merge into one file per program | ✅ | Entry 0013 redesign: 1 654 → 236 files for SoSe 2026; flat {period}/{program}.md layout, max 2 levels deep. |
| Weekly GitHub Action — Campo | ✅ | Mondays 03:00 UTC: scrape, render, commit, build zip, cut a Release with the zip + JSON intermediates. |
| Monthly GitHub Action — FAU.de | ✅ | First of each month, 04:00 UTC; the next weekly run inlines the new content. |
| Full-depth catalogue walk | ⏸ | --max-depth 0 is technically supported but a single GH-Actions job hits the 6-h timeout. A daily-resume chain is the next available step if and when deeper data turns out to be needed. |
| Archive prior years | later | archive/{year-slug}/ per completed academic year. |
See docs/requirements.md §9 for the full plan and open items (token-counting library, course-association strategy, next semester's periodId).
Only public Campo data is in scope. Anything that needs a FAU login (personal schedule, grades, exam registrations, enrolments) is explicitly out — and out forever, because the corpus is shipped publicly on GitHub.
MIT © Andreas Maier.
The Campo data the corpus references (titles, schedules, descriptions) remains FAU's. Permalinks to Campo are preserved on every file so attribution and source-of-truth are one click away.