A dual-encoder, or two-tower, architecture uses separate encoders for search queries and candidate listings. It helps bridge the gap between a guest’s wording and a host’s description when exact-term matching misses relevant properties. section 4.2.6
The separation supports search at scale:
- Precompute embeddings for millions of listings; generate only the query embedding when a search arrives. section 4.2.6
- Optimize the query encoder for short, ambiguous phrases and the listing encoder for longer descriptions, amenities, and location details. section 4.2.6
Clicks after a search provide query–listing pairs that suggest relevance for training. A click is a behavioral signal, not an explicit relevance label. section 4.2.6
