Production AI systems rarely see clean text alone. Users upload screenshots, photos, PDFs, recordings, tables, and videos. A useful system has to combine those inputs, preserve the evidence behind its answer, and choose the right model or tool for each modality.

Multimodal applications support workflows such as screenshot troubleshooting, product search by photo, document analysis across text and diagrams, meeting summarization, video search, and voice agents with visual context.

This chapter covers the building blocks of multimodal systems: shared embeddings, cross-modal search, multimodal RAG, routing, and production architecture.