How It Works
CompKitchen answers a single question in two ways at once: what goes with this — and why? One answer comes from molecular food science. The other comes from 610K+ curated recipes. The interesting part is where they agree, and where they don't.
Two signals, one score
Shared flavor compounds
Every ingredient is a set of volatile molecules — the thing your nose actually detects. We compare two ingredients by how much of that molecular vocabulary they share, weighted by how rare each compound is. Two ingredients that share common compounds score lower than two that share the rare ones. (Technical name: IDF-weighted Jaccard.)
Recipe co-occurrence
We look at 610K+ curated recipes to see which ingredients chefs actually pair. This captures culinary wisdom that molecules alone miss — like why lemon goes with fish or cinnamon with apple. (Technical name: Normalized Pointwise Mutual Information.)
Combined
Default scoring is 50% molecular + 50% culinary, plus a synergy bonus when both signals agree. Switch to Science mode on the tools page for pure molecular scoring (70% compound Jaccard + 30% aroma-vector cosine similarity) — useful when you want chemistry, not tradition.
Novel pairings
The most interesting results are where the two signals disagree. When two ingredients share lots of flavor compounds but rarely appear in recipes together, we flag the combination as novel — chemistry says they should work, culinary tradition hasn't caught up yet. That's how "chocolate and blue cheese" or "strawberry and coriander" end up on famous tasting menus.
The tools
- Pair Finder — best matches for any ingredient, science or classic scoring.
- Novel Pairings — combinations chefs haven't explored yet.
- Bridge Finder — connect two ingredients that don't pair directly through a third.
- Substitutes — closest molecular stand-ins, optionally by culinary role (acid for acid, fat for fat).
- Mood Match — query by aroma descriptor: smoky, citrus, umami, floral.
- Pantry Mode — drop in what you have, see the best subsets and what's worth buying next.
- Compare — side-by-side compound diff for any two ingredients (/compare).
- Compound Explorer — click any compound to see which ingredients contain it, its odor threshold, and aroma class.
- Heatmap — visualize pairing strength across 31 popular ingredients at a glance (/heatmap).
- Recipe Browser — 610K+ curated recipes, full-text search + browse by ingredient or cuisine.
Who it’s for
Home cooks who want a reason behind a pairing. Professional chefs prototyping a new dish. CPG product developers exploring flavor territory. Food writers hunting for non-obvious combinations. Anyone who's ever stared at an ingredient and wondered what else.
The free tier gives you the top 5 ranked pairings and basic tool access. Pro ($12/mo or $99/yr) unlocks full ranked results, compound-level breakdowns with odor thresholds and OAV data, literature citations backing each pairing, CSV exports, saved pantries with custom notes, and opt-in search history. Enterprise tier adds JSON API access with higher rate limits for recipe platforms, product development tools, and AI applications.
Where the data comes from
The ingredient-compound graph is an aggregation of major public food-chemistry datasets, cross-validated and de-duplicated. The flavor compound count (5,240) is the union of:
FlavorDB 2.0
Ingredient-to-compound mapping maintained by IIIT-Delhi Computational Systems Lab. Our primary graph backbone.
FooDB
Comprehensive food chemistry database — adds ~28,000 compounds with structural, concentration, and CAS information.
Curated Recipe Corpus
610K+ recipes from OpenRecipes (CC BY 3.0), Fandom Recipes Wiki (CC BY-SA 4.0), and public-domain cookbook scans (Project Gutenberg, Internet Archive). Provides the culinary co-occurrence signal and full-text search corpus.
Ahn et al. Flavor Network
The seminal 2011 Scientific Reports paper dataset (1,530 ingredients, 1,107 compounds with CAS numbers). Licensed CC BY 4.0.
PubChem (NLM)
Chemical properties, molecular weights, and CAS validation for every compound in the graph. Public domain.
Flavornet (Cornell)
738 GC-olfactometry compounds with verified aroma descriptors — anchors aroma classification of structurally similar compounds.
ChemTastesDB
Curated taste classifications (sweet, bitter, umami, salty, sour, tasteless) for 2,944 compounds (Rojas et al., 2022). Used to validate our LLM-derived taste labels at 94% agreement. Licensed CC BY 4.0.
USDA Dr. Duke's Phytochemical DB
Phytochemistry of food plants from the Agricultural Research Service. Public domain.
Wikidata
Ingredient enrichment — taxonomy, common names, regional varieties. Surfaced on 91% of ingredient pages. CC0.
Odor threshold compendia
Selected detection thresholds for OAV calculations are taken from peer-reviewed compendia (including Volatile Compounds in Food, Maarse 1991) and treated as cited fact references. Each value is cross-checked against primary literature before use.
Cross-Validated Aroma Labels
Aroma family and chemical class assignments for 5,240 compounds, generated by Anthropic's Claude Sonnet and cross-validated at 94% agreement against ChemTastesDB's curated taste classifications.
Peer-reviewed GC-MS literature (44,648 papers)
An ongoing pipeline mines PubMed, Europe PMC, Semantic Scholar, and CORE for papers measuring flavor compounds in specific foods, extracts the data via LLM, validates CAS numbers against PubChem, and feeds verified results back into the graph. This is how the data stays current as new research publishes.
Our methodology paper (preprint)
Unsupervised Recovery of Food Taxonomy from Volatile Compound Profiles: A UMAP Analysis of 450 Ingredients
We show that ingredient identity can be recovered from volatile compound profiles alone, using UMAP dimensionality reduction. The clustering reproduces conventional culinary categories without supervision — evidence that our compound graph encodes real flavor structure, not artifacts of how data was collected.
Read the paper on ZenodoWhat it doesn't do
This is a flavor chemistry engine, not a recipe robot. It tells you what molecules two ingredients share and what recipes combine them — it doesn't tell you exact quantities, cooking times, or how texture and temperature interact. Molecular similarity is a hypothesis generator, not a guarantee: vanilla shares compounds with many things, but that doesn't make every combination delicious. Use the tools as a starting point, then cook.
Get in touch
A real person reads these. Email [email protected] for anything — questions about the data, a pairing that looks wrong, API access, licensing, or academic use. If something here contradicts what the engine actually does, that is a bug worth telling us about.