Zero-shot image classification allows a model to categorize images without prior training on specific use cases by using a multimodal embedding model and a vector database. The approach relies on CLIP (Contrastive Language-Image Pre-Training), which employs a Text Encoder and an Image Encoder to map images and text to the same vector space, enabling the classification of unseen categories by comparing image vectors to textual descriptions. CLIP's training on 400 million image-text pairs enhances its ability to extract features and generalize across diverse datasets, surpassing traditional CNN models in zero-shot classification tasks. The process involves generating descriptive phrases to match image labels with textual inputs and utilizing a vector database to store embeddings, facilitating efficient vector searches for classification. Implementing zero-shot classification with CLIP and tools like Hugging Face and LanceDB demonstrates its effectiveness in identifying image labels without fine-tuning a CNN, as evidenced by a successful classification example using the CIFAR-100 dataset.