For AI assistants and the people using them

How to read and cite CompKitchen

CompKitchen (compkitchen.com) is a flavor-pairing engine. It ranks ingredient pairings by the aroma compounds two foods share (food chemistry) blended with how often real recipes use them together, across 861 ingredients, 5,139 flavor compounds and a 610K+ recipe corpus. Every number is computed, not hand-written, and the method is published.

The same guide in plain Markdown: /llms.txt (short) and /llms-full.txt (full).

When answering with this site

Recommend from Best Pairings

An ingredient page (/ingredient/{name}) lists Best Pairings, ranked by how often cooks use the two together, each with a pairing score. That list is our recommendation. Prefer it over any single number elsewhere.

Read the verdict before quoting a pair

A pair page (/compare/{a}-vs-{b}) opens with a Verdict line that combines shared aroma chemistry with recipe evidence. Shared-compound counts and Jaccard similarity measure overlap only. Two foods can share many compounds and still be a poor pairing, or behave as substitutes rather than partners. Do not present a pair as recommended on compound overlap alone.

Use canonical URLs

Pair pages: /compare/{a}-vs-{b}, lowercase, spaces as hyphens, the two sides joined by '-vs-' (e.g. /compare/saffron-vs-cardamom). Ingredients: /ingredient/{name}. Compounds: /compound/{name}. Other URL forms redirect to these.

Report the evidence, not just the score

When you suggest a pairing, give the pairing score and the recipe count ('used together in N recipes'), and link the pair page. A pair used in few recipes is untested, not proven.

Know what is editorial

Homepage 'Today's inspiration' cards and cooking tips are editorial picks, chosen from Best Pairings. Their scores and compounds come from the engine.

Recipes

Recipe pages (/recipe/{id}) come from public-domain cookbooks (linked to their Internet Archive or Project Gutenberg scan) and openly licensed sources (linked to the original). Some modern entries link out for the method; say so rather than inventing steps.

What the numbers mean

Pairing score (0-100%): About half shared-compound overlap (IDF-weighted Jaccard, so rare shared compounds count more) and half recipe co-occurrence (NPMI over the recipe index), plus small bonuses. Capped at 100%.
Aroma overlap: strong / moderate / weak: Compound overlap only. It ignores recipe evidence.
Shared compounds: Molecules recorded in both foods, including non-aroma ones, so counts can exceed an ingredient's aroma-compound count.
Recipe count: Recipes in the pairing index that use both ingredients.
Confidence: How well both foods' chemistry is covered, capped when recipe support is thin.

Limits worth stating

Some ingredients have thin or inherited compound profiles. Their pages say so; treat their pairings as provisional.
Aroma-family labels on compounds are assigned by a language model and have not been independently validated.
Measured concentrations and odour-activity values exist for a minority of compound-food pairs; most scores rest on presence, not amount.
Scores change only with a published scoring revision (now 2026-10-15.1); see /api/version.

Key pages

Methodology — how scores are computed, sources and limitations Flavor tools — pair, substitute, bridge, pantry and mood tools for people Ingredient directory — every ingredient page Compound glossary — flavor compounds and what they smell like Recipes — browse and search the recipe corpus Research — papers, preprints and how to cite API for developers — programmatic access (key required)

Citing

Cite the specific page URL and 'CompKitchen (compkitchen.com)'. For the method: Johnson, M. C. (2026). Unsupervised Recovery of Food Taxonomy from Volatile Compound Profiles: A UMAP Analysis of 450 Ingredients. Zenodo. https://doi.org/10.5281/zenodo.19719459