Embeddings as a Retrieval Index

Embed users and items into the same vector space and serving collapses to one operation: items near this user vector. The approximate nearest-neighbor index doesn't just store the model's output — the index is the model. All the expensive structure lives in the geometry.

Embeddings do double duty in a recsys: as dense features fed to downstream models (the ranker consumes them like any other input), and as the retrieval index itself for candidate generation. That second role is what made million-item catalogs servable — precompute every item vector once, then per-request work is a millisecond-scale ANN lookup.