How Apple TV Ranks Search Results From Two Keystrokes (RecSys 2026)
Apple's search-as-you-type system gives every title two embeddings — a text encoder fine-tuned on co-engagement and an ID embedding learned from co-watching — and fuses per-keystroke similarities with boosted trees for +2.99% NDCG@10.
Where this fits in Mixpeek
Personalized, as-you-type search is a fusion problem: multiple similarity signals combined and reweighted per query. Mixpeek's retrievers compose stages and fuse signals (semantic, behavioral, metadata) — learned fusion of several embedding spaces, exactly what this Apple system does, is the pattern behind Mixpeek's fusion and learned-weight retriever stages for personalized ranking.
About this research
Search-as-you-type gives a ranker 1-3 characters of intent. Apple's production system gives every show two embeddings: a text encoder fine-tuned on co-engagement, and an ID embedding learned purely from co-watching. Your watch history becomes a user vector in both spaces and boosted trees fuse the similarities per keystroke: +2.99% NDCG@10 offline, confirmed in a three-week online A/B. Paper: arxiv.org/abs/2607.13493
Frequently asked questions
Put the research to work
Mixpeek turns video, images, audio, and documents in your object storage into searchable, timestamped results through one API — the retrieval stack these papers describe.
Search your own data, free