Ingredient prediction
Infer likely ingredients from dish names, descriptions, cuisine context, modifiers, and preparation language.
The Honeycomb model
Food intelligence starts with better data.
Our purpose-built model predicts ingredients, allergens, dietary factors, macros, and micronutrients from the imperfect menu data restaurants actually have.
INPUTSpicy tuna poke bowlAhi tuna, sushi rice, avocado, sesame, miso mayo…
ENRICHED OUTPUTAllergensFish · Sesame · SoyDietsPescatarianNutritionMacros + microsStatus✓ Ready
Built for food-service groups, delivery and ordering platforms, POS systems, and hospitality software.
SODEXONOONFOODEE
Our menu intelligence has also labelled catalogs for some of the world's largest food delivery platforms.One model, structured output
Predict what is inside an item,
not just what it is called.
Menu names and descriptions hide the context that guest-facing products need. Honeycomb turns that sparse text into structured food intelligence.
Allergens and diets
Predict allergen presence and dietary compatibility with food-specific reasoning built for real menu edge cases.
Nutrition enrichment
Return consistent macro and micronutrient structures that downstream teams can map into their own products.
The work behind the model
Accuracy was earned
one hard label at a time.
The useful data for food intelligence is rarely available off the shelf. We spent the hours curating and labelling it, then built custom evaluations around the failures that matter to restaurants and their guests.
Build the dataset the category was missing.
Real menu items, ingredient context, allergen rules, dietary nuance, and the messy exceptions that generic datasets leave out.
Measure the decisions that matter.
Custom evals test ingredients, allergens, diets, and nutrition separately so one aggregate score cannot hide a dangerous miss.
Turn every failure into a better model.
We inspect edge cases, improve the data and training loop, then rerun the same evals before a model earns its way into production.
Keep the intelligence lightweight.
Purpose-built inference makes it possible to label millions of items in minutes instead of waiting days for general compute-heavy models.
Integration surfaces
Meet your stack
where it works.
Use the REST API inside your application, move catalog work through the CLI, or let agents operate through MCP.
request.sh
curl -X POST https://api.honeycomb.ai/v1/label \ -H "Authorization: Bearer $TOKEN" \ -d '{"name":"Garden Greens Salad", ...}'
$ hc label ./seasonal-menu.csv \ --output enriched-menu.json ✓ menu labels ready
You: Update the spring menu and regenerate the allergen guide for every location. Honeycomb: Menu updated. All location artifacts are ready for review.Illustrative interface for first-version design
“
The best model is not always the biggest one. It is the model shaped by the right data, tested against the right failures, and light enough to use everywhere.
AshayaFounder, Honeycomb AI
Designed for catalog scale
Millions of items.
Minutes, not days.
Purpose-built, lightweight inference processes large catalogs consistently, preserves provenance, and returns structured results that downstream products can use immediately.
MillionsItems per catalogMinutesTo structured labels
Honeycomb model
IngredientsAllergensDietsNutrition
Bring your menu data
Let’s make it
useful everywhere.
Talk to us about your catalog, platform, and integration surface.