Searching across different kinds of content at once: finding the slide with a particular chart, the photograph matching a description, the moment in a recording where something was said. It requires representing all of it in a comparable way, which joint embeddings provide.
Technical view
Indexes text, images, audio segments and video frames in a shared or aligned embedding space, permitting cross-modal queries. Practical designs often combine modality-specific retrieval with fusion, and generate text descriptions of non-text content so that mature text retrieval machinery can be reused, which is frequently more effective than pure cross-modal embedding. Evaluation requires cross-modal relevance judgements, which are expensive to obtain.
Related
Part of AI Terminology Index, section: Modalities and Applications.