Latimal

Food delivery platforms have a discovery problem. Customers search "something spicy" or "cold coffee" and the keyword matcher returns nothing useful. Cuisine tags drift and miss across thousands of restaurants. Cart upsell prompts ignore what is already in the basket. And on top of all that, a single dish appears as "Murgh Makhani" on one POS system, "Butter Chicken" on Swiggy, and "बटर चिकन (Serves 2)" on Zomato, leaving catalogs full of silent duplicates.

Generic embedding models don't solve any of these well. They can't tell that "cold coffee" and "Iced Americano" should surface together in search results. They tag "Tonkotsu Ramen" as generically Asian rather than specifically Japanese. They have no idea that "**BUY 1 GET 1** Margherita Pizza [Medium, Cheesy Crust, Limited Time]" is just a pizza with marketing noise wrapped around it. They treat "Kadhai Chicken" and "Karahi Chiken" as unrelated strings. And they can't distinguish a dietary conflict ("Classic Burger" vs "Veggie Burger") from a true duplicate.

The underlying challenge is specific: multilingual, noisy, food-domain text with transliterations across scripts, regional naming conventions that vary by city, and promotional cruft injected by restaurant operators. General-purpose models do not handle this distribution. So platforms either build fragile rule systems or live with broken search and duplicate listings.

dish-embed is a food-domain AI API. It reads dish names the same way across spellings, transliterations, abbreviations, and 100+ languages, and it ignores the promotional noise restaurants wrap around them. Hindi-English transliteration, cross-script matching, and food-specific semantics work out of the box.

Separate retrievers handle search (graded relevance) and deduplication (binary same/different), because the two tasks have opposite failure modes. We evaluate against 25 internal benchmarks spanning Indian, global, cross-lingual, beverage, bakery, and category-specific scenarios.

The API covers semantic search, cuisine classification, cart upsell, cross-restaurant price comparison, menu deduplication, and menu health reports. It supports 100+ languages and all major world cuisines. Credit-based access at dish-embed.latimal.com.