Ranking and retrieval models

How this question breaks down

1. 1Two towersCore17 minEncode the user and the item with separate networks so the score is a dot product; that one constraint lets you embed every item ahead of time and find the best few among millions with a nearest-neighbour index.
2. 2Learning to rankCore15 minA ranker is judged on the order of a short list, not on each score alone. Pointwise, pairwise and listwise losses get progressively closer to that, and LambdaRank's trick is to weight each mistake by how much it hurts the top of the list.
3. 3ANNCore25 minExact nearest-neighbour search reads every vector; an ANN index reads a small fraction by partitioning the space, walking a graph, or compressing the vectors, so a Pantrio query touches about 0.2% of 6 million recipe vectors and lets a small share of the true neighbours slip past.

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