A training approach that adapts a generative language model to produce high-quality text embeddings by repurposing its existing knowledge without building from scratch.