Candidate Generation (Recall)

- Collaborative filtering: matrix factorization (ALS) or neural CF (NCF) - Content-based: item embeddings via sentence/transformer model, ANN search (Faiss / HNSW) - Graph-based: random walk on user-item bipartite graph (PinSage / LightGCN) - Trending / popular: non-personalized fallback - Multi-channel recall: combine 5-10 recall sources, each returns top-K (200-500) - Recall service: real-time embedding + ANN lookup, 50ms budget