munhim/semantic-product-search

Semantic Product Search Model

This model performs semantic product search using BERT embeddings and a dual-encoder neural network architecture.

Model Architecture

• Base Model: BERT-base-uncased for text embeddings
• Encoder: Dual-encoder architecture with separate query and product encoders
• Similarity Network: Multi-layer perceptron for relevance scoring
• Input Dimension: 768 (BERT embedding size)
• Hidden Dimensions: [512, 256, 128]
• Dropout: 0.3

Usage

See the load_and_run_frontend.py script for loading and using this model.

Files

• pytorch_model.bin: Model weights
• config.json: Model configuration
• tokenizer files: BERT tokenizer files
• product_catalog.parquet: Product catalog for search
• product_embeddings.npy: Precomputed product embeddings (optional)

Performance

Trained on Amazon Shopping Queries Dataset with the following metrics:

• NDCG@10: ~0.54
• MAP: ~0.54
• Precision@10: ~0.50
• Recall@10: ~0.54