Winterlicious 2026 Menu Explorer

A fast, filterable, mobile-friendly browser for every restaurant participating in Toronto's Winterlicious 2026 prix-fixe festival.

Live: https://asterling.github.io/WinterLicList/

Why this exists

The official Winterlicious site is slow, paginated, and hard to filter. This one loads every menu at once and lets you slice the list however you like:

• 🔎 Search restaurant names and dishes
• 🍱 Filter by cuisine, neighbourhood, lunch/dinner price
• 🌱 Veg / Vegan only toggle (only restaurants with qualifying items)
• ⭐ Michelin highlight
• ❤️ Favorites saved locally
• 🗺️ Map view powered by Leaflet + OpenStreetMap
• 🔗 Deep links — share any restaurant: /#r=canoe
• 📱 Mobile friendly

How it works

• winterlic.py — Python scraper hitting Toronto's open-data endpoint. Auto-detects whether the active campaign is Winter- or Summerlicious based on the calendar (override with LICIOUS_SEASON=Winter|Summer).
• menus-latest.json — the freshest snapshot, always the file the site loads first.
• menus-{Winterlicious,Summerlicious}-YYYY.json — per-campaign archives.
• season.json — {season, year, label, fetched_at}; drives the page title.
• index.html / styles.css / app.js — static frontend, no build step, no framework.

Open index.html over http:// (not file://) so the JSON fetch succeeds. The quickest local server:

python3 -m http.server 8000
# then open http://localhost:8000

Refreshing the data

python3 -m pip install requests
python3 winterlic.py

This rewrites winterlicious_menus_2026.json from the city's live endpoint.

A GitHub Action runs this on a schedule — see .github/workflows/refresh-data.yml.

AI enrichment (optional, local LLM)

Each restaurant in the JSON can be augmented with four AI-generated fields:

• ai.standouts — 3–5 menu items the LLM thinks are worth flagging, taken verbatim from the menu
• ai.vibe_tags — 3–5 short tags from a controlled list ("date night", "casual", "upscale", "groups", "trendy", "cozy", "celebration", "hidden gem", "neighbourhood favourite", …)
• ai.one_liner — a single-sentence pitch that replaces boilerplate city descriptions on the card and modal
• ai.dietary_summary — deterministic, not LLM: counts veg / vegan / GF items from the city's flags

The enrichment runs entirely locally via Ollama. No API keys, no per-query cost. Results are cached by content hash in enrichment-cache.json, so unchanged menus skip the LLM on subsequent runs.

One-time setup

1. Install Ollama — either the prebuilt installer from https://ollama.com/download, or build from source:

   git clone --depth=1 https://github.com/ollama/ollama
   cd ollama
   # macOS users on Xcode 15+ may need: -ldflags="-extldflags=-Wl,-ld_classic"
   go build -o ollama .

2. Start the Ollama server (leave it running in another terminal):

   ollama serve

3. Pull the default model (~4.7 GB on disk, ~5 GB RAM at runtime):

   ollama pull qwen2.5:7b

   Smaller alternatives if you're tight on RAM (lower quality):

   • qwen2.5:3b — ~2 GB
   • qwen2.5:1.5b — ~1 GB (often hallucinates dish names; not recommended)

Running it

# Fetch + enrich in one shot
python3 winterlic.py --enrich

# Just enrich the JSON you already have
python3 winterlic.py --no-fetch --enrich

# Pick a different model
python3 winterlic.py --enrich --model qwen2.5:3b

# First N restaurants only (handy when iterating on prompts)
python3 winterlic.py --enrich --limit 10

# Force re-enrichment ignoring the cache
python3 enrich.py menus-latest.json --force

Runtimes on Apple Silicon for the full 240-restaurant sweep:

Model	Steady-state	First-run total	Cached re-run
qwen2.5:7b	~5 s/restaurant	~20 min	seconds
qwen2.5:3b	~3 s/restaurant	~12 min	seconds
qwen2.5:1.5b	~3 s/restaurant	~12 min	seconds

The frontend reads ai.* fields if present and falls back gracefully if they're missing, so enrichment is always optional.

Data source

https://secure.toronto.ca/c3api_data/v2/DataAccess.svc/Licious/map_data — City of Toronto open data, no API key required.

License

MIT. Not affiliated with the City of Toronto or Winterlicious.