A map of 1,790 ingredients, placed by how the world cooks with them. Explore what sits beside what, find a stand-in for something you have run out of, and see where each of my cookbooks lives.
Each dot is one of 1,790 ingredients, and each tinted patch is a region of them. Two dots sit close when cooks use them the same way, so miso lands beside dashi and tarragon beside chervil. The positions fall out of 4.1 million recipes; I only named the regions.
Loading 1,790 ingredients…
Two answers for any ingredient: the swaps cooks already rely on, with quantities, and its nearest neighbours in flavour.
Loading 1,790 ingredients…
Lay a book’s ingredients over the map and its cuisine shows up as territory. Each cuisine also has a direction in the space, so a book’s lean toward it can be measured.
Loading 1,790 ingredients…
Epicure represents every ingredient as 300 numbers, learned so that ingredients met in similar company end up with similar numbers. It ships three versions. One learns only from which ingredients share recipes, one only from which share flavour compounds, and one blends the two. The neighbour lists let you switch between them.
The positions are Epicure’s own projection of those 300 numbers onto a plane. A projection keeps close things close but cannot keep every distance, so neighbours are always computed in the full space and then drawn on the map. A link that crosses half the map is a real neighbour the flat picture could not seat nearby. The tilt and the raised regions are only there to make the map easier to read: a region’s height means nothing, while a pin’s height shows the same quantity as its colour.
Dots that sit together are grouped into 18 regions by position alone, so each region is one unbroken patch. I named each after what it holds. The tints carry no meaning beyond keeping neighbours apart, as on a political map, and a region’s edge is a convenience: ingredients near a border belong a little to both sides.
The general substitutions come from the notes on ingredient pages of the Wikibooks Cookbook, restated in my own words with a link to each page. They cover about a hundred ingredients. Where that source says nothing stands in well, the finder says so too.
Epicure tags each ingredient with the cuisines it belongs to. A cuisine’s direction runs from the centre of all ingredients to the centre of that cuisine’s own. An ingredient’s leaning is how far it points along that direction, in standard deviations, and a book’s leaning is the average over its ingredients, weighted by how many recipes use each.
Every ingredient line in the library is matched against the atlas. Matching is by words: plurals are reduced, British and American names are paired, and the longest name wins, so coconut milk is never coconut plus milk. Only counts are kept. No recipe text is published here.
Where it bends. Over a third of the training recipes are Chinese or Korean, so East Asian ingredients are mapped in finer detail than South Asian or Latin American ones. Ingredient names were merged with the help of a language model, and rare ingredients have few recipes behind them. Distances between common ingredients are the most reliable.
Radzikowski and Chen, 2026. The three ingredient embeddings used throughout this page, trained on 4.14 million recipes from eleven sources in several languages. Released under CC BY 4.0.
The published vectors and the 1,790-name vocabulary. They are redistributed here as unit vectors rounded to eight bits, which shifts a cosine by about 0.002 on average.
The two-dimensional map coordinates and each ingredient's food group, diet flags and cuisine tag. MIT licensed.
The source of the known swaps: the substitution notes on its ingredient pages and its general substitutions table. Licensed CC BY-SA 4.0; the restated table on this site is offered under the same licence.
The same authors' earlier study showing that taste, texture, geography and processing can be read out of ingredient embeddings as directions.