SMARTER SEARCH AT AIRBNB: HOW EMBEDDING-BASED RETRIEVAL WORKS

Summary from the TLDR Web Dev newsletter

Airbnb improved its search system using Embedding-Based Retrieval, which represents both listings and search queries as vectors in a shared space to find relevant matches. The system used a two-tower model architecture, where one tower processes listing information and the other processes search query information to create embeddings that were compared to quickly narrow down results. Airbnb trained this model using contrastive learning, pushing similar queries and listings closer together while distancing dissimilar ones.