Nova 2 Multimodal Embeddings converts text, documents, images, video, and audio into unified numerical vectors for cross-modal RAG, semantic search, and classification with 8K token context.