5.4. Reflection Checkpoint

Key Takeaways

• Same backbone, different payload. Text analysis, multimodal understanding, speech, and image generation all ride the two-client Foundry pattern; what changes is the input and output.
• Text analysis via prompt: one generative model does sentiment, entities, keywords, and summarization. Ask for structured output and low temperature when a program consumes it.
• Multimodal understanding = image in, information out through one model. It's the opposite of image generation.
• Speech has two directions: recognition (speech-to-text, input) and synthesis (text-to-speech, output). Azure Speech in Foundry Tools provides both; multimodal models can take spoken prompts directly.
• Image generation = text in, image out. Prompt specificity drives quality. Don't confuse generating an image with analyzing one.

Connecting Forward

You've now covered generation, agents, and the text/speech/vision modalities. One major workload remains: pulling structured information out of unstructured content. Phase 6 covers Azure Content Understanding — extracting fields from documents, images, audio, and video — which is the last piece of the implementation domain and a brand-new emphasis on this exam.

Self-Check Questions

• For each, name the modality capability and the data direction: (a) read a receipt photo aloud to a user; (b) generate a banner image from a tagline; (c) tell whether a spoken customer call sounds frustrated.
• Explain why a single multimodal model call can answer "what's in this image?" but you'd reach for a different model capability to create an image.