The 300-Millisecond Impression Funnel
Every time a user opens a Meta app, an ad impression opportunity triggers a high-speed auction. In roughly 300 milliseconds, Meta Andromeda must evaluate millions of candidate ads and select the single best creative to present.
This retrieval process is divided into multiple stages: broad-phase retrieval, fine-phase scoring, and absolute ranking. Andromeda filters down candidates from millions to thousands, then hundreds, and finally the top-performing ad based on a combined value formula that weighs bid size, estimated action rates, and user value.
Candidate Generation Phase
During candidate generation, the engine filters the global ad set pool using coarse-grained models. It targets broad parameters like geographic presence, language, and initial age brackets to yield a subset of roughly 10,000 viable candidate ads for heavy scoring.
Multidimensional Embeddings and Deep Learning Matchers
Andromeda relies heavily on deep neural networks to produce continuous, dense vector spaces representing users and ad creative features. Rather than matching exact keywords, the system projects users and ads into a shared 256-dimensional embedding space.
Creative Feature Extraction and the GEM Integration
A major development in Andromeda is its integration with Meta's Generative Ads Model (GEM). GEM analyzes creative image, video, and audio assets to construct rich multi-modal feature vectors.
This means Andromeda can 'understand' that a video shows a high-tempo demonstration of kitchen knives, allowing the delivery engine to target cookery enthusiasts with extreme precision. Creative elements are no longer black boxes; they are parsed into multi-dimensional coordinates.