Two towers are not independent of neural networks. Most two tower models I encounter are neural networks for each tower with optionally a mixing towering/layer

Two towers are not independent of neural networks. Most two tower models I encounter are neural networks for each tower with optionally a mixing towering/layer after the two. And sometimes it's not two but even more towers depending on how your features are grouped. I would consider two towers pretty much a necessity for large corpus retrieval as step 1 in any rec system with many items and many requests. Stage 2 or later models can be heavier and be whatever you like.