6.4 Practice Questions: Computer Vision, NLP, and Generative AI
Key Takeaways
- Domains 3-5 (45-60% combined) reward precise service-to-task mapping for vision, NLP/speech, and generative AI.
- Computer vision: pre-built analysis = Azure AI Vision, text-in-images = OCR, your own classes = Custom Vision, faces = Face, structured form data = Document Intelligence.
- NLP/Speech: text analysis = Language, per-aspect sentiment = opinion mining, intent/entities = CLU, FAQ bot = custom question answering, audio = Speech, text translation = Translator.
- Generative AI: RAG fixes made-up facts, low temperature = consistent, high temperature = creative, few-shot = examples, chain-of-thought = step-by-step, content filters and Prompt Shields provide safety.
- Service renames are tested: LUIS->CLU, QnA Maker->Custom QA, Form Recognizer->Document Intelligence, Cognitive Search->AI Search, AI Studio->AI Foundry.
Test your knowledge of Domains 3, 4, and 5 with these questions covering Azure AI services for vision, language, speech, and generative AI. Together these domains make up 45-60% of the exam — computer vision (15-20%), NLP (15-20%), and generative AI (20-25%).
Domain 3: Computer Vision — Service Selection
| Need | Service / capability |
|---|---|
| Pre-built captioning, tagging, object detection | Azure AI Vision |
| Read text from images/photos | Azure AI Vision OCR (Read) |
| Train a model on YOUR images/classes | Azure AI Custom Vision |
| Detect/verify/identify faces | Azure AI Face (gated) |
| Extract key-value pairs, tables, and layout from forms | Azure AI Document Intelligence |
Watch the OCR vs. Document Intelligence distinction: basic OCR returns raw text, while Document Intelligence understands structure — key-value pairs ("Invoice Number: 12345"), tables, and selection marks. When a scenario provides a small labeled image set for niche classes, the answer is Custom Vision, not pre-built Vision.
Domain 4: NLP and Speech — Service Selection
| Need | Service / capability |
|---|---|
| Sentiment, key phrases, NER, PII, summarization | Azure AI Language (pre-configured) |
| Per-aspect sentiment (rooms vs. food vs. service) | Opinion mining (aspect-based sentiment) |
| Understand a user's intent and entities | CLU |
| FAQ bot from existing documents | Custom question answering |
| Transcribe or synthesize audio | Azure AI Speech (STT / TTS) |
| Translate written text or documents | Azure AI Translator |
| Translate live spoken audio | Azure AI Speech translation |
Two frequent traps: opinion mining (not basic sentiment) is the answer when the scenario asks for sentiment per aspect; and document translation that "preserves formatting" is Translator, not a generative model.
Domain 5: Generative AI — Key Decisions
| Symptom in the scenario | Best answer |
|---|---|
| Bot invents facts about your products | Implement RAG with Azure AI Search to ground answers |
| Need consistent, factual output | Low temperature (near 0) |
| Need creative, varied output | Higher temperature (0.7-1.0) |
| Provide example input-output pairs | Few-shot prompting |
| Ask the model to reason step by step | Chain-of-thought |
| Block harmful prompts/responses | Azure OpenAI content filters (on by default) |
| Stop jailbreak/injection attempts | Prompt Shields |
Remember the service renames, which the exam tests directly: LUIS -> CLU, QnA Maker -> Custom Question Answering, Form Recognizer -> Document Intelligence, Cognitive Search -> Azure AI Search, Cognitive Services -> Azure AI Services, Azure AI Studio -> Azure AI Foundry. If you see an old name, map it to the current one.
Worked Reasoning: Three Easy-to-Confuse Pairs
OCR vs. Document Intelligence. Both read text from images, but if the scenario needs structure — invoice fields, tables, key-value pairs — choose Document Intelligence. If it just needs the raw text off a sign or package, plain OCR (Azure AI Vision Read) is enough.
Custom Vision vs. pre-built Vision. Pre-built Azure AI Vision recognizes general objects and scenes out of the box. The moment a scenario provides your own labeled images for niche classes the model would not already know (specific species, specific products, specific defects), the answer is Custom Vision.
RAG vs. fine-tuning. When a bot must answer from changing or proprietary documents, choose RAG (Azure AI Search + Azure OpenAI). Fine-tuning changes the model's overall style or behavior and is not the way to inject specific, frequently updated facts. "Made-up product facts" -> RAG; "always respond in our brand voice" -> fine-tuning.
The One-Sentence Test for Service Selection
Before answering any Domain 3-5 question, finish this sentence: "The input is ___ and the required output is ___." Examples:
- Input = audio, output = text -> Azure AI Speech (speech-to-text).
- Input = text, output = the same text in another language -> Azure AI Translator.
- Input = image, output = structured fields -> Document Intelligence.
- Input = a question, output = an answer grounded in our docs -> RAG (Azure AI Search + Azure OpenAI).
This input/output framing cuts through most distractors, because the wrong services either take the wrong input (text vs. audio) or produce the wrong output (analysis vs. generation).
How to Use This Practice Set
Answer the full set before reading explanations, then mark each miss by service name and capability. For every wrong answer, write why the correct service wins and why the distractor service does not fit. These domains reward precise service-to-task mapping more than any other part of the exam.
Review Routine
Keep a three-column log: scenario clue, the correct service, and the distractor you almost chose. Retake missed questions the next day and mix them with new ones. Before answering, restate the input and the required output in your own words ("image in, structured fields out" -> Document Intelligence); that single habit resolves most service-selection questions.
Because Domains 3-5 together are the majority of the exam, invest your final review here. Build a one-page service map — vision services, language/speech services, and generative-AI services with their key capabilities — and recite which service solves each common scenario until the mapping is automatic. The exam rarely asks you to configure anything; it asks you to pick the right service, and a well-rehearsed service map makes that quick and reliable.
A logistics company wants to automatically read tracking numbers from packages using cameras on a conveyor belt. Which Azure AI service and capability should they use?
A pharmaceutical company needs to translate their drug interaction documentation from English into 30 languages while preserving the formatting of tables and chemical formulas. Which service should they use?
A hotel chain wants to analyze thousands of online reviews to understand what guests like and dislike about specific aspects of their stay (rooms, food, service, location). Which Azure AI Language feature provides the most detailed analysis?
A company built a chatbot using Azure OpenAI Service. Users report that the chatbot sometimes makes up information about company products that does not exist. What is the best solution?
Which of the following is the correct current name for the service previously known as "Azure Cognitive Search"?
A customer service team wants a chatbot that can answer employee questions about company benefits using their existing benefits handbook PDF. Which TWO Azure services should they use together?
What happens when a user submits a prompt containing hate speech to Azure OpenAI Service with default settings?
A developer wants the AI model to give consistent, factual answers with minimal variation. Which temperature setting should they use?
Which prompt engineering technique involves providing the model with example input-output pairs before presenting the actual task?
A wildlife conservation organization needs to identify specific endangered species in camera trap images. They have 500 labeled images of 10 species. Pre-built Azure AI Vision does not recognize these specific species. Which service should they use?
Which of the following best describes what Azure AI Document Intelligence does compared to basic OCR?
A company wants their speech recognition system to accurately transcribe medical terminology during doctor-patient consultations. Which approach should they take?
Which THREE features does Microsoft Copilot use to implement responsible AI? (Select three)
Select all that apply
Match each retired Azure service name to its current replacement:
Match each item on the left with the correct item on the right
You've completed this section
Continue exploring other exams