Visual search

How this question breaks down

1. 1EmbeddingsCore24 minTrain an image embedding in which a street photo of a jacket lands next to the seller's studio shots of the same jacket, using listing photos, search logs and a small human-labelled set, a classification-style metric-learning loss, and a vector small enough to store 52 million times.
2. 2IndexCore23 minStore 52 million photo vectors so that a query finds its nearest neighbours in a few tens of milliseconds, while 1.5 million photos arrive and 1.4 million leave every day: choose what to index, how to compress it, how to split it by category, and how to swap it when the model changes.
3. 3RankingCore22 minTurn a thousand visually close photos into a page a shopper can buy from: collapse duplicates across sellers, drop what's sold or not in their size, score the rest with visual, attribute and listing features trained on graded purchase signals, and show nothing rather than a page of wrong matches.

Work through the parts in order and you’ve covered the full design. Each part is one topic.