Refining machine learning models efficiently requires surfacing high-impact data for labeling, a process complicated by the challenge of navigating extensive unstructured data. Labelbox addresses this issue by offering natural language search within its Catalog, enabling users to instantly and accurately locate relevant data through well-crafted prompts. Effective prompt engineering, an increasingly valuable skill with the rise of large language models, involves using simple, specific language, incorporating visual cues, and iterating prompts to enhance search precision. Labelbox provides users with a system to refine prompts using positive and negative biases, score ranges based on cosine distance, and the ability to combine natural language searches with other filters. This approach not only streamlines data discovery but also integrates seamlessly with Labelbox's suite for enhanced search and classification, ultimately aiding in tasks like zero-shot learning and bulk classification, which expedite the labeling process by automatically applying annotations.