Study Notes

Appearance

AI-102 Exam Tips & Traps ​

💎 General Tips ​

  • Branding Matters: The exam heavily uses Microsoft AI Foundry (or just "Foundry tools"). If you see "Azure AI Studio," that's the older name—Foundry is the new standard.
  • Python/C# Focus: You don't need to be a pro coder, but know the common SDK patterns (e.g., DefaultAzureCredential, ProjectClient, ChatCompletionsOptions).
  • REST vs SDK: Most questions favor the SDK. Know the difference between a 202 async pattern (OCR) and a synchronous call.

🧊 Domain Specific Tips ​

Domain 1: Plan & Manage ​

  • Hub vs Project: Hub is for infrastructure/shared settings; Project is for your specific app/model testing.
  • PTU vs Standard: Choose Provisioned Throughput (PTU) for predictable latency and enterprise scale. Choose Standard (PAYG) for development and variable traffic.

Domain 2: Generative AI ​

  • RAG vs Fine-tuning:
    • Choose RAG for factual grounding on dynamic data.
    • Choose Fine-tuning for specific tone, complex output formats, or deep domain specialized knowledge (where RAG recall is poor).
  • Prompt Flow: Know the node types (LLM, Python, Prompt). Evaluation flows are for measuring Groundedness and Relevance.

Domain 3: Agentic Solutions ​

  • Tool Calling: The model doesn't "run" the function; it outputs a JSON call that your application code runs.
  • Code Interpreter: Use for math, sorting, or data analysis tasks that LLMs often hallucinate on.

Domain 4: Vision ​

  • OCR READ API: Always use the async pattern (Submit -> Check Status -> Get Results).
  • Video Indexer: Remember it extracts insights (faces, topics, sentiment), not just OCR.

Domain 5: NLP ​

  • Custom QA: Replaces QnA Maker. Know that you can import from URLs, PDF manuals, and even .chitchat files.
  • CLU Entities: Use Prebuilt entities (Email, Number) whenever possible instead of manual ones.

Domain 6: Knowledge Mining ​

  • Skillset Errors: If an enrichment fails, check the Indexer Execution History.
  • Vector Search: Requires Embeddings (e.g., text-embedding-3-small). Vectorization can be "integrated" (built-in) or manual.
  • Hybrid Search: Always the "best" answer for complex retrieval (Keyword + Vector + Semantic Ranking).

⚠️ Common Traps ​

  • Rate Limits: If you get 429 errors, the answer is usually "Implement exponential backoff" or "Increase TPM quota," not "Switch to a different model."
  • Content Safety: Content filtering happens at the resource level in Azure OpenAI. If the model ignores bad words, check your Content Safety filter severity.
  • Search Latency: If search is too slow, use HNSW (Hierarchical Navigable Small World) algorithm for vector indexing.