Design a Two-Tower Retriever With Cold-Start Handling

A rental marketplace has 5M active listings, about 8 percent of which are less than 7 days old and have almost no interaction history, and hundreds of millions of (user, listing) engagement events. Search is a ranker with rich cross features, but its candidate generator (popularity plus filters) misses relevant listings. You are asked to add an embedding-based retrieval stage with a 50 ms budget for producing ~500 candidates.

1. Sketch the two-tower architecture: inputs to each tower, the scoring function, the loss and where negatives come from.
2. Explain how serving works and why the structure of the model makes that possible. Estimate the index memory for 128-d float32 and int8 vectors.
3. Name three features you would like to use but cannot in a tower, and explain where they belong instead.
4. How do you make new listings retrievable on day one? Cover both training and serving.
5. How do you evaluate the stage offline and what would you check before an A/B test?