Incremental video search requires high-quality rating after every keystroke, the place intent is usually underspecified (e.g., 1–3 character prefixes). We current a personalization system for Apple TV search that mixes complementary semantic and collaborative alerts at rating time. Our strategy learns two merchandise embedding areas: (i) a text-based multilingual encoder (TextEmb) fine-tuned on co-engagement triplets through contrastive studying, and (ii) an ID-based collaborative embedding mannequin (IdEmb) skilled on interaction-derived positives. At serving time, we assemble consumer representations from latest watch historical past and inject text- and ID-based consumer–merchandise cosine similarities right into a pairwise XGBoost ranker. We consider the system with temporally held-out offline datasets and a three-week on-line managed experiment. Offline, for periods with consumer historical past, the customized ranker improves NDCG@10 by 2.99% and MRR by 3.30% over the non-personalized baseline. Crucially, slice analyses present that personalization is most wanted in incremental search, the place intent remains to be forming: on ambiguous prefix queries (1–3 characters), NDCG@10 raise is +8.63%, versus solely +1.46% on longer, extra totally specified queries. Customers with longer watch histories profit extra from personalization than newer customers: NDCG raise rises from +2.13% for customers with 1–5 historical past objects to +4.37% for customers with 51–100. This bigger raise happens although baseline relevance is decrease for long-history cohorts (NDCG@10 drops from 0.733 to 0.680), indicating that personalization provides probably the most worth the place default rating underperforms. On-line, therapy yields statistically vital beneficial properties of +1.14% tap-through fee and +1.23% conversion fee, with a 2.91% enchancment in converted-item rank place. We additional analyze protection–precision trade-offs between semantic and collaborative embeddings by means of ablations isolating every sign, and consider embedding high quality on a held-out corpus with LLM-judged similarity labels to cut back click on/publicity bias.
