- 01Catalog6 demo · ~10M real
- 02Retrieval5 demo · ~1K · ~30 ms
- 03Ranking~1K → top 100
- 04ResultsTop 100 · ~100 ms total
Candidates for ranking
Retrieval removes 1 irrelevant item; ranking places 3 of the 5 candidates into results.
Knit walking shoe
Rank #1
Cushioned trainer
Rank #2
City sneaker
Rank #3
Trail hiking shoe
Candidate
Canvas low-top
Candidate
Coffee maker
Filtered by retrieval
WHY RETRIEVAL
Don't run a pairwise model on all 10 million products
Score every pair10M comparisons
At 100K pair scores / sec~100 sec / search
Combine three retrieval routes~30 ms
1KeywordsExact terms + attributes
2Semantic SLMTwo-tower model
3Personalization DNNTwo-tower model
→
~1Kto ranking
Full search, end to end~100 ms total
HOW THE TWO-TOWER ROUTES WORK
Map query and product into the same space
Semantic SLM and Personalization DNN both use two towers: one encodes the query, the other each product. Similar query–product meanings land close together.
-
OfflineProducts→Item tower→
OnlineQuery→Query tower→
itemquery
same vector space
Two towers, one vector space
Product vectors are precomputed offline; each search creates a query vector.
-
ANN searches nearby groups
It probes likely clusters or graph neighbors, not all 10M vectors.
Further reading: two-tower retrieval and ANN · vector index trade-offs.