Two-stage video recommender

Design a two-stage candidate-generation + rerank recommender for a short-video app with 200M DAU and a 1B-video corpus, where each 'next videos' request must return a diverse, fresh slate in under 120ms p99. Stage one retrieves a few thousand candidates from a billion-item corpus; stage two reranks to a final ~20 with a heavier model. Cover both stages, how candidate generation stays unbiased and diverse, and how the two stages are kept aligned.