"Semantic Retrieval at Walmart"

Walmart's hybrid search combines neural and traditional retrieval to crack the tail query challenge.

Walmart developed a hybrid search system combining traditional inverted index with neural retrieval to handle millions of daily product searches, particularly improving tail query performance through efficient embedding-based semantic matching and practical deployment optimizations.

-----

https://arxiv.org/abs/2412.04637

🔍 Original Problem:

→ E-commerce product search faces unique challenges compared to web search, especially for tail queries with specific intent

→ Traditional text matching methods struggle with vocabulary mismatches and synonyms

→ Pure neural retrieval systems are limited by embedding size constraints and latency requirements

-----

🛠️ Solution in this Paper:

→ The system uses a two-tower BERT architecture to generate embeddings for queries and products

→ A novel negative sampling strategy combines product category matching and token matching to improve model training

→ Linear projection reduces embedding dimension from 768 to 256 while maintaining performance

→ The architecture merges results from both inverted index and neural retrieval before final ranking

-----

💡 Key Insights:

→ Product titles provide most signal for retrieval compared to descriptions

→ Freezing token embeddings during training improves model generalization

→ Hard negative sampling significantly boosts category recall by 20.47%

-----

📊 Results:

→ NDCG@10 improved by 2.84% for tail queries

→ Add-to-cart rate increased by 0.54%

→ Maintained low latency of 13ms for ANN service