recipe-search-sbert

Recipe Search SBERT

This model combines SBERT embeddings with a denoising autoencoder to create recipe embeddings optimized for semantic search in Portuguese recipes.

Model Description

The model processes recipe texts into embeddings using a two-stage approach:

• SBERT (distiluse-base-multilingual-cased-v2) generates initial 768-dimensional embeddings
• A denoising autoencoder reduces dimensionality to 128 while preserving recipe relationships

Intended Use

• Portuguese recipe semantic search
• Recipe similarity analysis
• Ingredient-based recipe recommendations

Training Data

The model trains on 15,000 recipes from TudoGostoso. Each recipe contains:

• Title
• Ingredient list
• Cooking instructions
• URL reference

Training Procedure

1. SBERT generates embeddings from concatenated recipe texts
2. Mean pooling creates single vector per recipe
3. Denoising autoencoder trains for 100 epochs with:
   • Adam optimizer (lr=1e-4)
   • MSE loss
   • Noise factor=0.1
   • Batch size=32

Evaluation Results

Embedding space evaluation via TSNE visualization shows:

• Clear clustering of similar recipes
• Sensible transitions between recipe types
• Preserved relationships after dimensionality reduction

Example Usage

from transformers import AutoTokenizer
from recipe_search_sbert import RecipeSearchEngine

# Initialize model
search_engine = RecipeSearchEngine()

# Search recipes
results = search_engine.search("frango assado")

Limitations

• Portuguese recipes only
• Single-cuisine dataset bias
• No support for cross-lingual search