5.4. Reflection Checkpoint

Key Takeaways

• Multimodal models handle most visual understanding via prompts — captioning, visual Q&A, chart reading — collapsing former "train a model" tasks into prompt-and-image.
• An image in a prompt is grounding-by-attachment; the same prompt levers (instructions, structured output, grounding) apply.
• Dedicated Vision services win for specialized/high-volume needs — OCR/Read, face detection, video/spatial analysis — and custom models for your own trained categories.
• The recurring trap is over-reaching for a custom model or classifier when a multimodal prompt or prebuilt capability already fits — and the inverse, ad-hoc prompting where precise OCR is needed.

Connecting Forward

Phase 6 covers text analysis, which mirrors this domain's lesson exactly: an LLM-first reframing where generative models with structured output now do much of what dedicated NLP services used to, while the Azure Language service still earns its place for high-volume, PII, and specialized cases.

Self-Check Questions

• For each task, name the right tool and justify: (a) generate alt-text for a blog image, (b) extract text from 10,000 scanned invoices with confidence scores, (c) classify factory photos into your three proprietary defect types.
• Explain the AI-102-to-AI-103 shift in computer vision in one sentence, and give one task that did not move to multimodal prompting.