MvLFood

Food options around Maria-von-Linden-Strasse.

Small installable web app for quick lunch decisions near Maria-von-Linden-Strasse.

Run locally

npm run dev

Then open http://localhost:5173.

Install on iOS and Android

This is a Progressive Web App. On Android, open the URL in Chrome and use install/add-to-home-screen. On iOS, open it in Safari and use Share > Add to Home Screen.

Data approach

The app only reads data/menu.json. That keeps the phone app fast, offline-friendly, and independent of cross-origin rules or PDF parsing.

The normalized feed has:

• sources for Hungry Elk, MPH, and future food trucks
• days[].items[] for all visible food options
• diet values: vegan, vegetarian, meat, fish, or unknown
• allergens using the source codes printed on the plans

Food trucks should be added as another source with "kind": "truck" and normal items for the days they appear. The UI will automatically show the source badge.

Source logos live in data/images/logos/. The Max Planck mark was looked up from Wikimedia Commons' Max Planck Forschung SVG, and the Hungry Elk mark was derived from a provided screenshot by extracting the elk line art into a small transparent PNG.

Spicetripping should only be added to the feed for the days explicitly mentioned in the weekly email. Save the email body as text and the attached menu image into data/images/trucks/, then run:

npm run parse:spicetripping -- --email=/path/to/email.txt --image=data/images/trucks/spicetripping-YYYY-MM-DD.png

The helper extracts the service date, time, and location from the email and emits normalized JSON to paste into data/menu.json. The menu image can be attached to truck items with the optional image field.

For Max-Planck-Haus, parse the HTML Speiseplan table instead of the PDF text so category and diet come from the table headers:

npm run parse:mph

Update mechanism

The feed is refreshed by .github/workflows/update-menu.yml:

• It runs every Monday at 08:17 Europe/Berlin time, with a 10:47 backup run to reduce the chance of a missed GitHub scheduled event. It can also be started manually from GitHub Actions.
• It runs npm run update:menu, then npm run validate.
• If data/menu.json changed, the workflow commits the new feed back to the repository.

The updater currently owns the sources with stable public menu pages:

• Hungry Elk: discovers the current PDF through the Stollsteimer WordPress REST API and parses the positioned PDF text.
• Max-Planck-Haus: parses the public HTML Speiseplan table.

If OPENAI_API_KEY is available, npm run update:menu also sends the generated Hungry Elk and MPH items through an LLM normalization pass. The LLM only rewrites app-facing title, description, and diet; dates, sources, categories, prices, allergens, and opening hours stay under deterministic code and validation.

For GitHub Actions, add an OPENAI_API_KEY repository secret. You can optionally set an OPENAI_MODEL repository variable; otherwise the workflow uses gpt-5-mini. To force a deterministic-only local run:

MENU_LLM_NORMALIZE=0 npm run update:menu

Recurring non-dated entries such as Yellow Donkey are carried forward from the existing JSON for matching weekdays. Manually entered dated truck visits, such as Spicetripping emails, are preserved only when their openingHours.date falls inside the generated week.

For local testing:

npm run update:menu
npm run validate

npm run update:source-text is still available as a debugging helper when you want to inspect raw PDF text.

Open questions

• Which food trucks should be included, and where is their most reliable schedule/menu source?
• Should allergy filtering hide unknown items, or keep them visible with an ask badge?
• Which additional food-truck feeds have reliable public sources that can be parsed automatically?