Menuvision

Pipeline Components

The AI agent creates these scripts:

DATA CONTRACT (Critical)

All three pipeline stages share this exact JSON schema. The AI agent MUST use these field names — any deviation breaks the pipeline.

menu_data.json Schema

{
  "restaurant": {
    "name": "Restaurant Name (if visible)",
    "cuisine": "cuisine type (Chinese, Indian, Austrian, Japanese, etc.)",
    "tagline": "any subtitle or tagline"
  },
  "sections": [
    {
      "title": "Section Name (in primary language)",
      "title_secondary": "Section name in secondary language (if present, else empty string)",
      "category": "food or drink",
      "note": "Any section note (e.g. 'served with rice', 'Mon-Fri 11-15h')",
      "items": [
        {
          "code": "M1",
          "name": "Dish Name (primary language)",
          "name_secondary": "Name in secondary language (if present)",
          "description": "Brief description (primary language)",
          "description_secondary": "Description in secondary language (if present)",
          "price": "12,90",
          "price_prefix": "",
          "allergens": "A C F",
          "dietary": ["vegan", "spicy"],
          "variants": []
        }
      ]
    }
  ],
  "allergen_legend": {
    "A": "Gluten",
    "B": "Crustaceans"
  },
  "metadata": {
    "languages": ["German", "English"],
    "currency": "EUR"
  }
}

Field Reference

EXTRACTION PROMPT

Send this exact prompt to Gemini. It defines the schema AND the extraction rules. Do not paraphrase it.

You are a restaurant menu data extractor. Analyze this menu content and extract ALL items into structured JSON.

Return this exact JSON structure:
{
  "restaurant": {
    "name": "Restaurant Name (if visible)",
    "cuisine": "cuisine type (Chinese, Indian, Austrian, Japanese, etc.)",
    "tagline": "any subtitle or tagline"
  },
  "sections": [
    {
      "title": "Section Name (in primary language)",
      "title_secondary": "Section name in secondary language (if present, else empty string)",
      "category": "food or drink",
      "note": "Any section note (e.g. 'served with rice', 'Mon-Fri 11-15h')",
      "items": [
        {
          "code": "M1",
          "name": "Dish Name (primary language)",
          "name_secondary": "Name in secondary language (if present)",
          "description": "Brief description (primary language)",
          "description_secondary": "Description in secondary language (if present)",
          "price": "12,90",
          "price_prefix": "",
          "allergens": "A C F",
          "dietary": ["vegan", "spicy"],
          "variants": []
        }
      ]
    }
  ],
  "allergen_legend": {
    "A": "Gluten",
    "B": "Crustaceans"
  },
  "metadata": {
    "languages": ["German", "English"],
    "currency": "EUR"
  }
}

CRITICAL RULES:
1. Extract EVERY item. Do not skip ANY dish, drink, or menu entry.
2. Preserve original item codes/numbers if present (M1, K2, S3, etc.). If none exist, generate sequential codes per section (e.g. A1, A2 for appetizers, M1, M2 for mains).
3. Extract prices EXACTLY as written (preserve comma/period format).
4. If an item has a price prefix like "ab" (starting from), capture it in "price_prefix".
5. If an item has multiple size/quantity variants (e.g. 6 Stk / 12 Stk / 18 Stk at different prices), use the "variants" array:
   [{"label": "6 Stk", "price": "8,90"}, {"label": "12 Stk", "price": "15,90"}]
   In this case, set the main "price" to the smallest variant's price.
6. Capture allergen codes exactly as shown (letters, numbers, or symbols).
7. If an allergen legend is visible anywhere, include it in "allergen_legend".
8. Identify dietary flags from descriptions/icons: vegan, vegetarian, spicy, gluten-free, halal, kosher.
9. If the menu is bilingual, capture BOTH languages. Put the primary/dominant language in name/description and the secondary in name_secondary/description_secondary.
10. For set menus or lunch specials with a fixed price covering multiple choices, create a section with note explaining the format, and list each choice as an item.
11. Classify each section as "food" or "drink".
12. For drinks, still extract name, price, and any size variants.

Return ONLY valid JSON. No markdown fences, no explanatory text.

Vision Prompt Variant

For image-based inputs (screenshots, PDF pages, photos), prepend a context line before the base prompt:

EXTRACTION_PROMPT_VISION = (
    "You are a restaurant menu data extractor. "
    "This is a photo/scan of a restaurant menu page.\n\n"
    "Return this exact JSON structure:"
    + EXTRACTION_PROMPT.split("Return this exact JSON structure:")[1]
)

Then each input type adds its own prefix:

GEMINI API CONFIGURATION

import os
from google import genai

client = genai.Client(api_key=os.environ["GOOGLE_API_KEY"])

def gemini_config():
    return genai.types.GenerateContentConfig(
        max_output_tokens=65536,          # 64K — needed for large menus
        response_mime_type="application/json",  # JSON mode — critical
    )

# Model: gemini-2.5-flash (default)
response = client.models.generate_content(
    model="gemini-2.5-flash",
    contents=prompt_text,    # or [image, prompt_text] for vision
    config=gemini_config(),
)

# ALWAYS check for truncation
if response.candidates[0].finish_reason.name == "MAX_TOKENS":
    print("WARNING: Response truncated. Menu may be incomplete.")

IMAGE PROMPT TEMPLATE

Use this exact function. It produces the casual phone-photo aesthetic that makes menus look authentic.

def build_food_prompt(name: str, description: str, cuisine: str = "") -> str:
    cuisine_context = f" {cuisine}" if cuisine else ""
    food_desc = f"{name}"
    if description and description != name:
        food_desc += f" ({description})"

    return (
        f"A photo of {food_desc} at a{cuisine_context} restaurant. "
        f"Taken casually with a phone from across the table at a 45-degree angle. "
        f"The plate sits on a dark wooden table and takes up only 30% of the frame. "
        f"Lots of visible table surface around the plate. Chopsticks, napkins, "
        f"a glass of water, and small side dishes scattered naturally nearby. "
        f"Blurred restaurant interior in the background — other diners, pendant lights, "
        f"wooden chairs visible but out of focus. Warm ambient lighting. "
        f"NOT a close-up. NOT professional food photography. "
        f"It looks like someone quickly snapped a photo before eating."
    )

IMAGE GENERATION API CALLS

Gemini 2.5 Flash Image

import os, io
from PIL import Image
from google import genai

client = genai.Client(api_key=os.environ["GOOGLE_API_KEY"])

def generate_gemini(client, name, description, output_path, cuisine=""):
    prompt = build_food_prompt(name, description, cuisine)

    response = client.models.generate_content(
        model="gemini-2.5-flash-image",       # NOT gemini-2.5-flash (that's text-only)
        contents=prompt,
        config=genai.types.GenerateContentConfig(
            response_modalities=["TEXT", "IMAGE"],  # critical — requests image output
        ),
    )

    # Extract generated image from response parts
    for part in response.candidates[0].content.parts:
        if part.inline_data is not None:
            img = Image.open(io.BytesIO(part.inline_data.data)).convert("RGB")
            # Center-crop to square, resize to 800x800
            w, h = img.size
            side = min(w, h)
            left = (w - side) // 2
            top = (h - side) // 2
            img = img.crop((left, top, left + side, top + side))
            img = img.resize((800, 800), Image.LANCZOS)
            img.save(str(output_path), "JPEG", quality=82)
            return
    raise RuntimeError("No image in Gemini response")

Skip drinks

Only generate images for category == "food" sections. Drinks get a text-only list in the HTML output.

MULTILINGUAL / CJK HANDLING

Menus can be in ANY language. The pipeline handles this through bilingual fields and smart prompt routing.

Extraction (all languages)

• name / description = primary language (whatever the menu is mostly written in)
• name_secondary / description_secondary = secondary language (if bilingual)
• Works for: German/English, Chinese/English, Japanese/English, Thai/English, Arabic/English, Korean/English, etc.

Image Generation (CJK-safe prompting)

CJK characters produce bad image prompts. Before calling build_food_prompt(), swap to the Latin name:

def prepare_for_image_gen(name, name_secondary, description):
    """Use Latin-script name for image prompts. CJK → use secondary name."""
    display_name = name
    if name_secondary:
        if any(ord(c) > 0x2E80 for c in name):  # CJK/Hangul/Kana detection
            display_name = name_secondary
            description = description or name
        else:
            description = description or name_secondary
    return display_name, description

Unicode ranges covered by ord(c) > 0x2E80:

• CJK Unified Ideographs (Chinese characters)
• Hiragana / Katakana (Japanese)
• Hangul (Korean)
• CJK Compatibility, Radicals, Extensions

HTML Output (all scripts)

• name renders as the large display text
• name_secondary renders below it in smaller text
• Both use Google Fonts with CJK fallback (Noto Sans SC, Noto Sans JP, Noto Sans KR)

FILE NAMING CONVENTIONS

Auto-derivation

All filenames are derived from the restaurant name or source URL:

stem = "shoyu"  # derived from URL domain, PDF filename, or restaurant name
data_file = f"menu_data_{stem}.json"
images_dir = Path(f"images/{stem}")
html_file = f"{restaurant_name}_Menu.html"  # e.g. "Shoyu_Menu.html"

Image files

images/{restaurant_stem}/{code}.jpg

# restaurant_stem = data filename minus "menu_data_" prefix
# Example: menu_data_shoyu.json → images/shoyu/M1.jpg

Image path matching (in build step)

Returns POSIX-style string paths with ./ prefix for cross-platform HTML compatibility:

def find_image(code: str, images_dir: Path):
    """Return relative POSIX path string to image, or None."""
    if not images_dir.is_dir():
        return None
    rel = images_dir.as_posix()
    if not rel.startswith("./"):
        rel = "./" + rel
    # 1. Exact match
    for ext in ("jpg", "jpeg", "webp", "png"):
        candidate = images_dir / f"{code}.{ext}"
        if candidate.exists():
            return f"{rel}/{code}.{ext}"
    # 2. Case-insensitive fallback
    for f in images_dir.iterdir():
        if f.stem.lower() == code.lower() and f.suffix.lower() in (".jpg", ".jpeg", ".webp", ".png"):
            return f"{rel}/{f.name}"
    return None

Output HTML

{RestaurantName}_Menu.html    # CSS/JS inline, images as relative file paths

Image rendering (build step)

The build script uses find_image() to resolve each food item's photo, falling back to a gradient SVG placeholder when no image exists:

import base64
import html as html_mod

GRADIENT_COLORS = [
    ("#c41e3a", "#8b0000"), ("#ff6b6b", "#ee5a24"), ("#fdcb6e", "#e17055"),
    ("#00b894", "#00cec9"), ("#6c5ce7", "#a29bfe"), ("#e17055", "#d63031"),
    ("#00cec9", "#0984e3"), ("#fab1a0", "#e17055"), ("#e8a87c", "#d4956b"),
    ("#fd79a8", "#e84393"),
]

def make_placeholder_svg(code: str, name: str, secondary: str = "") -> str:
    """Generate a base64-encoded SVG placeholder when no image exists."""
    idx = hash(code) % len(GRADIENT_COLORS)
    c1, c2 = GRADIENT_COLORS[idx]
    display = html_mod.escape(secondary[:12] if secondary else name[:12])
    svg = f'''<svg xmlns="http://www.w3.org/2000/svg" width="220" height="180" viewBox="0 0 220 180">
  <defs><linearGradient id="g" x1="0%" y1="0%" x2="100%" y2="100%">
    <stop offset="0%" style="stop-color:{c1}"/>
    <stop offset="100%" style="stop-color:{c2}"/>
  </linearGradient></defs>
  <rect width="220" height="180" fill="url(#g)" rx="12"/>
  <text x="110" y="75" text-anchor="middle" fill="rgba(255,255,255,0.25)" font-size="56" font-family="serif">{html_mod.escape(code)}</text>
  <text x="110" y="120" text-anchor="middle" fill="white" font-size="26" font-family="serif" opacity="0.9">{display}</text>
  <text x="110" y="148" text-anchor="middle" fill="rgba(255,255,255,0.6)" font-size="11" font-family="sans-serif">{html_mod.escape(name[:30])}</text>
</svg>'''
    b64 = base64.b64encode(svg.encode("utf-8")).decode("ascii")
    return f"data:image/svg+xml;base64,{b64}"


def image_tag(code: str, name: str, secondary: str, images_dir: Path, portable: bool = False) -> str:
    """Return <img> tag — real image OR gradient SVG placeholder.
    If portable=True, embed the real image as base64 data URI for single-file output."""
    real = find_image(code, images_dir)
    if real:
        if portable:
            img_path = images_dir.parent / real  # resolve relative path
            with open(img_path, "rb") as f:
                b64 = base64.b64encode(f.read()).decode("ascii")
            return f'<img src="data:image/jpeg;base64,{b64}" alt="{html_mod.escape(name)}">'
        return f'<img src="{html_mod.escape(real)}" alt="{html_mod.escape(name)}" loading="lazy">'
    else:
        src = make_placeholder_svg(code, name, secondary)
        return f'<img src="{src}" alt="{html_mod.escape(name)}">'

Output Modes

The HTML builder supports two output modes controlled by a --portable flag:

Portable mode embeds all food images as base64 data URIs directly in the HTML. File sizes are larger (~4-6MB for an 80-item menu) but the output is a single file that works everywhere with zero hosting setup. This is the default when no GITHUB_* environment variables are set.

Deployable mode uses relative image paths and requires the HTML file and images/ directory to be hosted together. Use this when publishing to GitHub Pages or any static hosting service.

ROBUSTNESS PATTERNS

Retry Logic

All Gemini API calls should retry on transient failures:

import time

def call_with_retry(fn, *args, max_retries=3, **kwargs):
    """Retry API calls with exponential backoff."""
    for attempt in range(max_retries):
        try:
            return fn(*args, **kwargs)
        except Exception as e:
            if attempt == max_retries - 1:
                raise
            wait = 2 ** attempt
            print(f"  Retry {attempt + 1}/{max_retries} in {wait}s: {e}")
            time.sleep(wait)

JSON Response Parsing

Gemini sometimes wraps JSON in markdown fences or produces trailing commas. Parse defensively — try raw parse first, apply trailing comma fix only as last resort (unconditional fix can corrupt valid JSON strings containing ,] patterns):

import re, json

def parse_gemini_json(raw: str) -> dict:
    """Parse JSON from Gemini, handling markdown fences and quirks."""
    text = raw.strip()
    # Strip markdown code fences
    if text.startswith("```"):
        text = re.sub(r"^```\w*\n?", "", text)
        text = re.sub(r"\n?```$", "", text)
    text = text.strip()
    # Try direct parse first
    try:
        return json.loads(text)
    except json.JSONDecodeError:
        pass
    # Try extracting JSON object from surrounding text
    match = re.search(r"\{.*\}", text, re.DOTALL)
    if match:
        candidate = match.group(0)
        try:
            return json.loads(candidate)
        except json.JSONDecodeError:
            pass
        # Fix trailing commas and retry
        candidate = re.sub(r",\s*([\]}])", r"\1", candidate)
        try:
            return json.loads(candidate)
        except json.JSONDecodeError:
            pass
    # Last resort: fix trailing commas on original
    text = re.sub(r",\s*([\]}])", r"\1", text)
    return json.loads(text)

Post-Processing

After extraction, run these cleanups:

def generate_codes(data: dict) -> dict:
    """Ensure every item has a unique code. Generates sequential codes per section
    if items have empty/missing codes (e.g. A1, A2 for appetizers, M1, M2 for mains)."""
    # ... assign prefix by section title, increment counter per section
    return data

def normalize_prices(data: dict) -> dict:
    """Normalize price formats: numeric → string, strip currency symbols,
    preserve comma/period format as-is."""
    # ... convert float/int to string, strip €/$, etc.
    return data

CURRENCY_MAP

Maps ISO currency codes to display symbols for the HTML output:

CURRENCY_MAP = {
    "EUR": "€", "USD": "$", "GBP": "£", "CHF": "CHF ",
    "JPY": "¥", "CNY": "¥", "INR": "₹", "AUD": "A$",
    "CAD": "C$", "SEK": "kr ", "NOK": "kr ", "DKK": "kr ",
    "THB": "฿", "KRW": "₩", "HKD": "HK$", "SGD": "S$",
    "CZK": "Kč ", "HUF": "Ft ", "PLN": "zł ", "TRY": "₺",
}

EXTRACTION DETAILS

HTML URLs

1. Fetch page with requests
2. Check text density to detect static vs JS-rendered: density = len(soup.get_text(strip=True)) / len(raw_html)
3. Density override: If 5+ price patterns found (r"[$€£¥₹CHF]\s*\d+[.,]\d{2}|\d+[.,]\d{2}\s*[$€£¥₹]"), force density to 1.0 (treat as static)
4. Static (density >= 0.02): Clean HTML, send text to Gemini 2.5 Flash (JSON mode)
5. JS-rendered (density < 0.02, e.g. Wix, Framer): Screenshot with Playwright, send to Gemini Vision
6. Screenshot height cap: If screenshot > 6000px tall, resize proportionally to fit
7. Large menus (>12k chars text): Chunked extraction, merge like PDF multi-page. Deduplicate by tracking seen_codes = set() across chunks — for each item in each chunk's sections, skip if item["code"] already in seen_codes. Only append sections that still have items after dedup.

PDF Files

1. Convert each page to image via PyMuPDF (200 DPI)
2. Send each page image to Gemini Vision
3. Merge results across pages (deduplicate items by code)

Photos

1. Load image directly
2. Resize if >10MB
3. Send to Gemini Vision

HTML OUTPUT FEATURES

• 3-column Instagram-style grid (9:16 portrait tiles)
• Gradient text overlay with name + secondary language + price
• Tap-to-select with green checkmark
• Receipt/bill on Selection tab with +/- quantity controls
• Category pill navigation with scroll sync
• Drinks section below grid with currency-prefixed prices
• Allergen legend
• Currency converter — minimalist button in header (e.g. € pill) that cycles or opens a picker for: EUR, USD, AUD, CAD, GBP. Converts all displayed prices client-side using snapshot exchange rates embedded at build time. Updates grid overlays, receipt totals, drink prices, and variant prices. Source currency comes from metadata.currency.
• Fully responsive, dark mode
• All CSS/JS inline, images via relative file paths (./images/{stem}/{code}.jpg), only Google Fonts external
• Gradient SVG placeholders for missing images (inline base64 SVG, not raster)
• CJK font loading via Google Fonts link tag: family=Noto+Sans+SC:wght@400;700&family=Noto+Sans+JP:wght@400;700&family=Noto+Sans+KR:wght@400;700
• CSS font-family stack: primary font, then 'Noto Sans SC', 'Noto Sans JP', 'Noto Sans KR', sans-serif

Currency Converter

A minimalist currency toggle built into the HTML output. All client-side, no API calls at runtime.

Implementation:

• The build script embeds a RATES object with snapshot exchange rates (base: USD) at build time
• Source currency is read from metadata.currency in the JSON data
• All prices are stored in data-price attributes as numeric values (not raw strings like "12,90")
• A small pill button in the header shows the current currency symbol (e.g. €)
• Tapping opens a mini-picker or cycles through: EUR (€), USD ($), GBP (£), AUD (A$), CAD (C$)
• On currency change, JavaScript converts all data-price values and updates displayed text
• Receipt totals in the Selection tab also convert via convertPrice() using SOURCE_CURRENCY and currentCurrency
• Variant prices also update
• Selected currency persists in localStorage

Price parsing helper (build-time — converts string prices to numeric for data-price attributes):

import re

def _parse_price_numeric(price: str) -> str:
    """Parse price string to numeric float for data-price attribute."""
    matches = re.findall(r"(\d+[.,]\d+)", price)
    if matches:
        return str(float(matches[-1].replace(",", ".")))
    return "0"

# Usage in HTML template:
# <div class="price" data-price="{_parse_price_numeric(item['price'])}">€12,90</div>

// Snapshot rates embedded at build time (base: USD)
const RATES = { EUR: 0.92, USD: 1.00, GBP: 0.79, AUD: 1.54, CAD: 1.36 };
const SYMBOLS = { EUR: "€", USD: "$", GBP: "£", AUD: "A$", CAD: "C$" };
const SOURCE_CURRENCY = "EUR";  // from metadata.currency

function convertPrice(amount, fromCurrency, toCurrency) {
    const inUSD = amount / RATES[fromCurrency];
    return inUSD * RATES[toCurrency];
}

// Applied to: grid overlay prices, drink list prices, variant prices,
// AND receipt/selection tab totals (all elements with data-price attribute)

The build script should fetch current rates at build time (or use reasonable defaults if offline). Prices display with 2 decimal places in the target currency, using the target locale's format.

Branding Customization

--name "Restaurant Name"     # Header brand text
--tagline "Cuisine · City"   # Subtitle
--accent "#ff6b00"           # Primary color (pills, active tab, drink prices)
--bg "#0a0a0a"               # Background color

COST SUMMARY

Drinks are not image-generated (text-only list), so actual cost depends on food-to-drink ratio.

DEPENDENCIES

Requires Python 3.9+.

Required:

• google-genai (extraction + image generation)
• Pillow (image processing)

For HTML URLs:

• requests (HTTP fetching)
• beautifulsoup4 (HTML parsing)

For JS-rendered sites:

• playwright (headless browser screenshots)

For PDF files:

• PyMuPDF (PDF to image conversion)

pip install google-genai Pillow requests beautifulsoup4 PyMuPDF
pip install playwright && playwright install chromium

ENVIRONMENT VARIABLES

• GOOGLE_API_KEY — Required for extraction and image generation
• GITHUB_PAT — Required for GitHub Pages publishing
• GITHUB_OWNER — Your GitHub username (default: reads from git config)
• GITHUB_REPO — Your GitHub Pages repo name (default: menus)

PUBLISHING

Default: Portable HTML (no setup)

When no GITHUB_* environment variables are set, the pipeline generates a self-contained HTML file with base64-embedded images. Users can:

• Open the file directly in any browser
• Email it or share via any file-sharing service
• Upload to any static host (Netlify Drop, Vercel, GitHub Pages, S3)

No hosting setup, no API keys beyond GOOGLE_API_KEY, no git configuration needed.

Optional: GitHub Pages (requires setup)

For users who want a persistent gallery with multiple menus:

1. Create a GitHub repo for your menus (e.g. your-username/menus)
2. Enable GitHub Pages on the main branch
3. Set environment variables (must be accessible to the Python process):

export GITHUB_PAT="your-personal-access-token"   # Required — used for git push auth
export GITHUB_OWNER="your-username"               # Required — YOUR GitHub username
export GITHUB_REPO="menus"                        # Optional — defaults to "menus"

Important: publish_menu.py MUST read GITHUB_OWNER and GITHUB_REPO from environment variables. Never hardcode a specific user's repo. The generated code should construct the repo URL dynamically:

owner = os.environ["GITHUB_OWNER"]
repo = os.environ.get("GITHUB_REPO", "menus")
GITHUB_REPO = f"{owner}/{repo}"
GITHUB_PAGES_BASE = f"https://{owner}.github.io/{repo}"

Publish

python publish_menu.py Restaurant_Menu.html --name "Restaurant" --tagline "Cuisine · City" --cuisine Type

Gallery: https://<your-username>.github.io/<repo>/

How publishing works

publish_menu.py clones the menus repo to a temp directory on native filesystem (git clone --depth=1), copies files there, commits, and pushes. This avoids all NTFS bind mount permission issues that occur when operating directly on mounted volumes in Docker containers.

Key implementation details:

1. git clone --depth=1 to a tempfile.mkdtemp() directory (native FS, proper POSIX permissions)
2. Copies HTML + images using shutil.copy() (not copy2 — avoids os.chmod() EPERM on NTFS)
3. find_image_dirs regex uses [^/"]+ (not [a-z_]+) to match Unicode chars in image dir names
4. Writes .meta_ JSON sidecar for gallery metadata
5. Rebuilds gallery index.html
6. Authenticates push via GITHUB_PAT env var embedded in the clone URL
7. Temp directory is cleaned up after push
8. MENUS_REPO_DIR (bind mount path) is only used for --list read-only queries

EXTERNAL ENDPOINTS

No analytics, telemetry, or tracking. No data is sent to any endpoint beyond those listed above.

SECURITY & PRIVACY

• API keys: GOOGLE_API_KEY is read from environment variables, never hardcoded or logged
• GitHub PAT: Used only for authenticated pushes to the user's own repo; never transmitted elsewhere
• Restaurant data: Menu content is sent to the Gemini API for processing. No data is stored server-side beyond Google's standard API retention
• Generated images: Stored locally in images/ directory. When published, uploaded only to the user's own GitHub Pages repo
• No telemetry: The pipeline collects no analytics, metrics, or usage data
• Local-first: All processing happens locally except Gemini API calls. The HTML output and images remain on the user's machine unless they explicitly publish

KNOWN LIMITATIONS

• Tabbed Wix menus: Only first visible tab extracted
• Google Maps photo URLs: Not supported (use direct image files)
• Very large menus (300+ items): May need manual chunk review