Researchers have developed a method to refine embedding models for zero-shot search and classification tasks using large language models (LLMs) as guides. This approach allows embeddings to adapt in real-time to the target task by leveraging feedback from a generative LLM on a small set of documents. The refinement process enables embedding models to better capture the nuances of a user query, leading to improved performance on challenging tasks. This technique has significant implications for applications where embedding models are used, such as information retrieval and text classification1. The ability to adapt embeddings in real-time can enhance the accuracy and effectiveness of these systems. So what matters to practitioners is that this approach can be used to improve the performance of embedding models in a variety of applications, making them more effective and efficient.