Designing a Two-Tower Retrieval Model
Problem statement
How would you build a two-tower retriever for a large recommender? Separate user and item networks produce unit-length vectors, and a dot product with temperature scaling scores each pair. Training uses other items in the batch as negatives, while serving retrieves candidates for a heavier ranker. Explain tower separation, training and serving.
Show hint
Ask why the two towers must not see each other's inputs.
Relevant roles: ML Engineer, Applied Scientist, AI Architect and Deep Learning Engineer.