closest_mode
ClaudeReturns the emergent GMM regions that best describe one ingredient. Each result includes a human-labelled mode, cosine similarity to its centroid, dominant cuisine and food group, and representative members. An optional property restricts the search to one axis family.
compare_on_axis
ClaudeCompares two ingredients on one named embedding axis, such as sweetness or a cuisine direction. Returns both projections, the difference from ingredient A to ingredient B, percentile labels, and the axis p10/p90 range. Valid axis names are available from list_targets with kind='direction'.
cultural_profile
ClaudeScores one ingredient against eight macro-region cuisine directions: Japanese, East Asian, Southeast Asian, South Asian, Latin American, Mediterranean, Eastern European, and Western Atlantic. Returns ranked cosine similarities with p10/p90 range context for cuisine-focused questions.
find_pairings
ClaudeExplores complementary ingredients for one or more seed ingredients. It is suited to open-ended pairing questions and recipe or dish design. Returns clustered flavour directions, each seed's strongest secondary connections, and bridge ingredients shared across clusters. Category penalties promote variety, and optional vegan or vegetarian filters remove incompatible suggestions.
flavour_correlations
ClaudeReports the strongest relationships between named axes in the global flavour space. Positive cosine values indicate aligned axes and negative values indicate opposing axes. Results can be limited and filtered by minimum absolute correlation to keep the response concise.
ingredient_on_factor
ClaudeProjects one ingredient onto one emergent ICA factor. Returns the signed projection, its percentile across all 1,790 ingredients, and the labels and anchoring ingredients for both poles. Positive values point toward pole A and negative values toward pole B; list_factors maps factor indices to named axes.
list_factors
ClaudeLists the 20 emergent ICA flavour factors available to ingredient_on_factor and pareto_navigate. Each summary includes the factor index, named axis, pole labels, and coherence ratings. Set include_examples=true to add themes and anchoring ingredients, and use min_coherence to filter the catalogue.
list_targets
ClaudeLists valid targets for morph and named axes for compare_on_axis. Directions include cuisines, sensory descriptors, nutrients, NOVA, and diet; modes are emergent GMM regions. Results are paginated by kind and include a concise angle_deg guide for morph.
morph
ClaudeRotates a seed ingredient toward a named direction, emergent mode, or second ingredient on the embedding unit sphere, then returns the nearest ingredients to the transformed vector. This supports explicitly directed transformations such as making an ingredient sweeter or more aligned with a cuisine. list_targets provides valid directions and modes. angle_deg ranges from 0 (unchanged) to 90 (fully orthogonal to the seed).
neighbors
ClaudeReturns the nearest ingredients to one seed by cosine similarity in the 300-dimensional embedding. The result is a ranked similarity list with no clustering or dietary filtering, suited to direct substitution and single-ingredient similarity questions.
pairing_score
ClaudeMeasures the affinity between two specific ingredients. Returns their 300-dimensional cosine similarity and a percentile label relative to ingredient pairs in the corpus. Higher values indicate a stronger learned pairing; typical scores fall between 0.0 and 0.6.
pareto_navigate
ClaudeComputes Pareto frontiers that balance similarity to a seed ingredient against movement toward one or more labelled flavour poles. It supports trade-off questions such as finding ingredients that remain rice-like while becoming more aligned with another culinary property. A specific factor and side can be supplied, or relevant poles can be selected automatically.
where_on_atlas
ClaudeLocates one ingredient on the precomputed 2D UMAP atlas. Returns its coordinate, cuisine and food-group labels, and nearest visual neighbours with 2D distances. The result describes spatial context rather than a named model region.
