Latimal Menu Intelligence

A model that only understands food

Food-domain AI for search, classification, cart upsell, matching, and dedup. The model reads dish names the same way across spellings, transliterations, and 100+ languages, treating “Margherita Pizza”, “Pizza Margherita”, “마르게리타 피자”, and “マルゲリータピザ” as one dish. The dedup demo below runs live; the rest of the toolkit is in the gallery further down.

Trained for food. Here is the proof.

Full benchmarks

Dedup F1 on the food benchmark suite. Latimal vs an off-the-shelf open-source stack (bge-m3 + bge-reranker-v2-m3).

Cross-lingual dedupSame dish across writing systems · F1

Latimal

0.83

bge-m3 + bge-reranker-v2-m3

0.26

Indian menu dedupCross-restaurant duplicate detection · F1

Latimal

0.92

bge-m3 + bge-reranker-v2-m3

0.75

Global cross-restaurant dedupDuplicate detection across world cuisines · F1

Latimal

0.82

bge-m3 + bge-reranker-v2-m3

0.80

Latimal: current production stack, benchmarked on 2026-05-22. Competitor: bge-m3 + off-the-shelf bge-reranker-v2-m3. Both run on the identical 26-benchmark snapshot.

Cross-script deduplication

Cluster duplicate menu items across spellings, transliterations, languages, and writing systems.

3601 msAPI

Example menus

Menu items, one per line15 items / 30 max

Paste your own menu above. The playground runs the first 30 items; the API itself handles up to 2,000 per call.

Nothing you paste in the playground is stored.

15 items

Verification is not set up here, so live runs are paused. You can still browse the examples.

15 items→4 dishes→11 duplicates removed

Seven scripts, one dish

Margherita Pizza written in Latin, Japanese, Korean, Arabic, Cyrillic, Chinese, and Thai, plus a typo, a promo tag, a portion tag, and a discount tag, all collapse into one cluster. Greek Salad, Buffalo Chicken Wings, and Chocolate Brownie stay separate. Keyword matching catches only the Latin variants and never crosses a writing system.

Canonical
ENMargherita Pizza
6 scripts
- ENPizza Margherita
  0.969
- ENMargherita Pizza (Large)
  1.000portion tag
- EN**NEW** Margherita Pizza
  1.000promo stripped
- ENMargerita Pizza
  0.838typo / spelling
- ENMargherita Pizza 50% OFF
  0.916promo stripped
- 日マルゲリータピザ
  0.565transliteration
- 한마르게리타 피자
  0.567transliteration
- عبيتزا مارغريتا
  0.620transliteration
- ??Пицца Маргарита
  0.597
- 中玛格丽特披萨
  0.536transliteration
- ไทพิซซ่ามาร์เกอริต้า
  0.568transliteration

server 3601 ms

How other tools cluster the same menu

The exact items you just ran, clustered three ways. Latimal pulls the variants into one dish; the others do not.

Latimal

15 → 4 dishes, 11 duplicates found

ENMargherita Pizza

ENPizza Margherita
ENMargherita Pizza (Large)
EN**NEW** Margherita Pizza
ENMargerita Pizza
ENMargherita Pizza 50% OFF
日マルゲリータピザ
한마르게리타 피자
عبيتزا مارغريتا
??Пицца Маргарита
中玛格丽特披萨
ไทพิซซ่ามาร์เกอริต้า

3 items left unmerged

bge-m3 + bge-reranker-v2-m3

15 → 13 dishes, 2 duplicates found

ENMargherita Pizza

ENPizza Margherita
ENMargerita Pizza

12 items left unmerged

Keyword search (SQL ILIKE)

15 → 12 dishes, 3 duplicates found

ENMargherita Pizza

ENMargherita Pizza (Large)
EN**NEW** Margherita Pizza
ENMargherita Pizza 50% OFF

11 items left unmerged

Methodology

Latimal. The live result from this playground run: Latimal's food-domain dedup, union-find clustering at cosine 0.85.
bge-m3 + bge-reranker-v2-m3. Off-the-shelf open-source stack: bge-m3 embeddings re-scored by bge-reranker-v2-m3, then union-find clustering at cosine 0.85 (the same threshold Latimal uses). Computed offline on these exact items.
Keyword search (SQL ILIKE). Two items group only when one is a case-insensitive substring of the other, the SQL `ILIKE '%term%'` pattern. No model, no embeddings. This is what menu deduplication looks like without semantic understanding.
The comparison is shown on curated examples only. The baseline columns are pre-computed; the Latimal column is the live result from your run above.

Precision check

Merging duplicates is only half the job. The model also has to refuse to merge dishes that merely look alike. These three pairs run live against the match endpoint.

Spelling variant
ENSpaghetti BologneseENSpaghetti Bolognaise
Same dishas expected
cosine
1.000
reranker
0.970
Merged as one dish: two accepted spellings of the same Italian classic.
Similar but distinct
ENPad ThaiENPad See Ew
Differentas expected
cosine
0.703
reranker
0.099
Kept apart: two different Thai stir-fried noodle dishes, not duplicates.
Different cuisines
ENMargherita PizzaENChicken Shawarma
Differentas expected
cosine
0.131
reranker
n/a
Kept apart: an Italian pizza and a Middle Eastern wrap share nothing.

Compare your own two items

Use this in your app

The snippet below carries the exact menu you just ran. Public Latimal API, swap in your key.

curl -X POST https://dish-embed.latimal.com/dedup \
  -H "X-API-Key: $LATIMAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "items": [
    "Margherita Pizza",
    "Pizza Margherita",
    "Margherita Pizza (Large)",
    "**NEW** Margherita Pizza",
    "Margerita Pizza",
    "Margherita Pizza 50% OFF",
    "マルゲリータピザ",
    "마르게리타 피자",
    "بيتزا مارغريتا",
    "Пицца Маргарита",
    "玛格丽特披萨",
    "พิซซ่ามาร์เกอริต้า",
    "Greek Salad",
    "Buffalo Chicken Wings",
    "Chocolate Brownie"
  ],
  "cosine_threshold": 0.85
}'

Run this on your full catalog

What you saw is a 15-item demo. The API clusters 2,000 items per call. Get a free key with 5,000 credits, or send the founder your messiest menu.

Get a key, 5,000 free credits Talk to the founder

The full menu intelligence toolkit

Semantic search, cuisine classification, cart upsell, and menu health, all powered by the same food-domain intelligence behind the dedup demo above. Open any tool to run it live.

Customers search by intent, not your exact menu wording.

query: “something spicy”

Tagging cuisine by hand does not scale across a global catalog.

Margherita PizzaItalianTacos al PastorMexicanTonkotsu RamenJapaneseMasala DosaSouth Indian

Report

Messy menus leak revenue through noise, dupes, and stray SKUs.

2 promo-noise items
4 duplicate clusters
2 non-food SKUs

Upsell

Generic add-on prompts ignore what is already in the cart.

cart: Tonkotsu Ramen + Chicken Gyoza

1Green Tea

2Ramune Soda

3Matcha Ice Cream