What are multimodal embeddings, and how are they used for cross-modal search?

Multimodal embeddings map different data types into one shared vector space where semantic similarity means proximity — a photo of a beach, the word 'beach', and a clip of waves all land near each other. Trained with contrastive objectives (CLIP for image-text, similar models for audio/video), they unlock cross-modal search: embed a text query and retrieve matching images, or embed an image and find related text, using ordinary nearest-neighbour search in a vector database. That's how 'search my photos by describing them' works. The catch is that alignment across modalities is imperfect — the spaces aren't perfectly fused — so retrieval quality varies and should be evaluated on labelled cross-modal pairs. Hybrid signals (captions, metadata) often help recover what pure embedding search misses.