1.3. Reflection Checkpoint

You now hold the two mental models that organize everything ahead: the five-layer anatomy of an AI solution, and the control-versus-abstraction axis for choosing services. Before moving into the heaviest exam domain, lock in the takeaways.

Key Takeaways

• Treat AI-200 as a distributed-systems exam, not a data science exam: the model is a dependency, and every question targets the system around it — data, containers, messaging, security, observability.
• Read every scenario for its constraint keyword ("kubectl," "scale to zero," "guaranteed delivery," "minimize overhead") — each keyword eliminates all but one service on the control-abstraction axis.
• Embeddings turn meaning into coordinates; vector search is nearest-neighbor geometry; ANN indexes trade a little recall for a lot of speed. This one idea underlies Cosmos DB vector search, pgvector, and Redis alike.
• Classify payloads before picking transport or storage: commands are messages (Service Bus), facts are events (Event Grid), access-granting values are secrets (Key Vault), behavior-shaping values are configuration (App Configuration).

Connecting Forward

Phase 2 dives into the exam's largest domain: the data layer. We'll take the vector-search principle from 1.1.2 and implement it three ways — in Cosmos DB for NoSQL, in PostgreSQL with pgvector, and in Azure Managed Redis — alongside each engine's operational mechanics: request units, indexing policies, change feeds, connection pooling, and caching patterns.

Self-Check Questions

• A scenario shows an API calling a model synchronously and timing out under load. Which layer of the anatomy is missing, and which two Azure services would you introduce to fix it?
• Why can an ANN-indexed vector search return different results than an exact search over the same data — and why is that usually an acceptable trade?