Two-Tower Model
A mainstream deep learning model used in the retrieval stage of search and recommendation systems. It consists of two independent encoders—a user tower and an item tower—that generate separate vector representations, enabling efficient offline indexing and online approximate nearest neighbor (ANN) retrieval.
Timeline (last 90 days)
In e-commerce scenarios, the Bi-encoder model is commonly used, with item-side vectors precomputed offline and stored in a database, while query-side vectors are generated online in real time.
The two-tower model uses a user tower and an item tower to generate embeddings separately, pre-computes item vectors offline and builds an index, and only needs to compute the query vector online before performing ANN retrieval.
The bi-encoder model encodes queries and documents into independent vectors respectively; document vectors can be pre-computed offline and indexed, enabling fast retrieval but unable to model token-level interactions.
The bi-encoder model can pre-compute document vectors offline and build an index for fast retrieval, but the two texts cannot see each other during the encoding stage, making it impossible to model fine-grained word-level interactions.
Cross-encoders jointly model the full content interaction between a resume and a job description through the attention mechanism, capturing fine-grained signals lost by dual-tower embedding models.
All Facts (5)
In e-commerce scenarios, the Bi-encoder model is commonly used, with item-side vectors precomputed offline and stored in a database, while query-side vectors are generated online in real time.
50%UnverifiedThe two-tower model uses a user tower and an item tower to generate embeddings separately, pre-computes item vectors offline and builds an index, and only needs to compute the query vector online before performing ANN retrieval.
50%UnverifiedThe bi-encoder model encodes queries and documents into independent vectors respectively; document vectors can be pre-computed offline and indexed, enabling fast retrieval but unable to model token-level interactions.
50%UnverifiedThe bi-encoder model can pre-compute document vectors offline and build an index for fast retrieval, but the two texts cannot see each other during the encoding stage, making it impossible to model fine-grained word-level interactions.
50%UnverifiedCross-encoders jointly model the full content interaction between a resume and a job description through the attention mechanism, capturing fine-grained signals lost by dual-tower embedding models.
50%