RecipeEZ

A Static, Ingredient-First Recipe Search and Visualization Engine

Overview • Architecture • Features • Research • Documentation

---

A purely data-driven interface for recipe retrieval.

Overview

Modern recipe websites are optimized for search engines and long-form narrative content rather than for cooking. RecipEZ is the opposite: a distraction-free, browser-rendered galaxy where the recipe data is the interface.

RecipEZ works as an inverse recipe search engine. Instead of searching for a specific dish, users enter ingredients they already have, receive a matching recipe seed, and explore its five semantically nearest recipes in a 3D point cloud.

The application ships without a live backend. Model vectors, UMAP coordinates, and K-NN neighbors are precomputed offline and loaded as static JSON.

Key Features

• Ingredient-first search: Enter available ingredients to find a contextual recipe seed, then inspect its precomputed neighbors.
• Four-model ablation: Toggle between combined SBERT, ingredient-only SBERT, name-only SBERT, and TF-IDF retrieval spaces.
• Model hover details: Each model button shows a short technical description on hover.
• Zero-backend infrastructure: Host the generated files with any static web server.
• 30,000-recipe galaxy: Explore a WebGL point cloud with interactive recipe details, ontology filters, and nearest-neighbor links.
• Reproducible research: Regenerate model metrics and comparison figures without rerunning embedding or UMAP training.

Architecture & Pipeline

Offline pipeline

1. Sample and clean recipes from the source dataset.
2. Build four feature representations:
   • MiniLM-L6-v2 embeddings for combined text
   • MiniLM-L6-v2 embeddings for ingredients
   • MiniLM-L6-v2 embeddings for recipe names
   • TF-IDF followed by 100-component SVD
3. Project each vector space to 3D with cosine-based UMAP.
4. Precompute five nearest neighbors in each original retrieval space.
5. Export static JSON bundles for the frontend.

Runtime frontend

1. Load one selected model bundle.
2. Render recipe coordinates as a Three.js point cloud.
3. Search locally by name or ingredient.
4. Open a recipe to display ingredients, instructions, and its five model-specific neighbors.

Project Structure

RecipEZ/
├── web/                  # Static Three.js frontend and generated model payloads
├── pipeline/             # Offline model builder and evaluation script
├── docs/                 # Architecture, pipeline, evaluation, and sequence diagrams
├── research/             # Reproducible metrics, comparison report, and figures
├── data/raw/             # Ignored raw dataset location
└── old/                  # Archived legacy apps, docs, tests, and presentations

Research

The active evaluation compares all four retrieval modes over the same 30,000 recipes, 21 ontology clusters, and 8,436 unique ingredient strings.

Model	Tag purity	Ingredient Jaccard	3D overlap
TF-IDF	0.5995	0.1143	0.1248
SBERT Combined	0.5406	0.1182	0.1421
SBERT Ingredients	0.3704	0.1785	0.1031
SBERT Names	0.4553	0.0969	0.1533

The result is not a single winner. TF-IDF best matches the current tag ontology, ingredient-only SBERT best preserves literal ingredients, and the SBERT modes offer concept-oriented alternatives. Most pairwise neighbor overlaps are below 0.06, so switching models meaningfully changes retrieval.

The browser starts in the Statistical (TF-IDF) mode by default, and it is the first model in the switcher.

Useful references:

• Model behavior radar
• UMAP projection gallery
• Model neighbor-overlap heatmap
• Cluster tag-purity heatmap
• Cluster ingredient-overlap heatmap
• Neighbor tag purity
• Ingredient Jaccard similarity
• Semantic vs geometry overlap
• PCA vs UMAP comparison

See Model comparison research for the complete report.

Documentation

• Documentation index
• Architecture
• Build pipeline
• Model evaluation
• Sequence diagrams
• Dataset and generated artifacts
• Model comparison research
• Legacy archive map

Quick Start

1. Install dependencies

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt

2. Provide the raw dataset

Download Food.com Recipes and Reviews and place RAW_recipes.csv at:

data/raw/RAW_recipes.csv

3. Generate the four static model bundles

python pipeline/build_models.py

This writes four 30,000-recipe JSON files to web/model_data/:

• sbert_all_data.json
• sbert_ingredients_data.json
• sbert_names_data.json
• tfidf_data.json

4. Serve the static frontend

python -m http.server 8000

Open:

http://127.0.0.1:8000/web/

5. Reproduce the research metrics

python pipeline/evaluate_models.py

This writes research/metrics.json and eight generated figures under research/figures/. The evaluation reads the model bundles and does not rerun SBERT, TF-IDF, SVD, or UMAP.

Model Modes

Mode	Browser label	Feature input	Vector space
TF-IDF	Statistical	Name, ingredients, and tags	TF-IDF followed by 100-component SVD
SBERT Combined	Combined	Name, ingredients, and tags	384D MiniLM-L6-v2 embeddings
SBERT Ingredients	Ingredients	Ingredient list	384D MiniLM-L6-v2 embeddings
SBERT Names	Names	Recipe title	384D MiniLM-L6-v2 embeddings

License

This project source is licensed under GPL-3.0. Dataset use is subject to the terms of the source dataset provider.