Multimodal joint embedding alignment

Design a multimodal embedding system for a marketplace where a user can search the catalog by uploading a photo OR typing text, and both must hit the same product index ('find me products that look like this image' and 'red leather ankle boots' against one shared store). You have 200M products, each with images and text descriptions. Design how you produce embeddings so images and text live in a comparable joint space, how you index for both query modalities, and how you keep the two modalities aligned as either model evolves.