6.1. LLM-First Text Analysis
💡 First Principle: Most text-analysis tasks are really "read this text and return a structured judgment" — which is exactly what a generative model with structured output does. The reframing: instead of calling a sentiment endpoint, you prompt a model to classify sentiment and return JSON. The skill is knowing when that's the right call versus when the dedicated service wins.
Why care: AI-103 explicitly emphasizes LLM-first text analysis. A single capable model with a well-designed prompt can classify, extract entities, detect sentiment, and summarize — reducing the number of services to integrate. But "can" isn't "should" for every case, and the exam rewards knowing the boundary.
⚠️ Common Misconception: "Sentiment, classification, and key-phrase tasks still require the dedicated Language service." Often a generative model with structured output replaces those endpoints — though the Language service still fits high-volume, cost-sensitive, PII, and specialized scenarios (text analytics for health). Treating the dedicated service as mandatory for every NLP task is the outdated default.
