Two-Tower Architecture

The previous chapter covered how embedding-based retrieval works at serving time: encode queries and items into vectors, search an ANN index, return candidates.

But the quality of those candidates depends entirely on how well the encoders are trained.

So the real question is: how do you train encoders that actually capture relevance?

This is where the two-tower architecture comes in. It’s the standard approach used at companies like YouTube, Google, and LinkedIn, and it powers most large-scale retrieval systems in production today.