How do retrieval + ranking pipelines use unsupervised embeddings?

They use unsupervised embeddings for candidate retrieval and learned ranking models for refinement.

Modern search systems operate in two stages: retrieval and ranking. Unsupervised embeddings (from contrastive learning or transformers) are used to perform fast approximate nearest neighbor search. Retrieved candidates are then passed through a ranking model that refines relevance using deeper interaction features. This hybrid architecture balances scalability and precision.