At inference, the target encoder is typically not needed; instead, the trained context encoder serves as a feature extractor that can be used for classification, retrieval, similarity search, or downstream supervised fine-tuning in a representation extraction application.