This final chapter consolidates everything you have learned and provides a structured preparation plan for the AI-102 certification exam. Use this as your go-to reference during the last days before your exam.
This domain covers the foundational aspects of working with Azure AI services. You should be comfortable with provisioning both single-service and multi-service resources, configuring authentication via keys and managed identities, securing services with virtual networks and private endpoints, and managing access through RBAC. Monitoring is a key focus: know how to configure diagnostic settings, interpret metrics like transactions, latency, and error rates, and set up alerts for anomalies. Cost management questions test your understanding of Free vs Standard tiers, commitment tier discounts, and budget alert configuration. Container deployment scenarios ask about which services support containers, how to configure billing and EULA parameters, and deployment targets like ACI, AKS, and IoT Edge. CI/CD pipeline questions focus on Azure DevOps and GitHub Actions for deploying AI infrastructure and models. Responsible AI principles — fairness, reliability, privacy, inclusiveness, transparency, and accountability — are tested throughout.
For Computer Vision, know the Analyze Image API features (tags, captions, objects, brands, landmarks, celebrities, adult content) and the Read API for OCR (printed and handwritten text, async operation pattern, result polling). Custom Vision requires understanding the difference between image classification (one tag per image) and object detection (bounding boxes), how to label images, train models via portal or SDK, and interpret evaluation metrics (precision, recall, mAP). Face API covers detection vs recognition, PersonGroup workflow for identification, verification endpoints, and attribute extraction (age, emotion, glasses). Video Indexer questions focus on insight extraction (faces, OCR, speech, sentiment, labels, keyframes, people tracking, scene detection) and customization options (brands, language models). Spatial Analysis involves container deployment for presence detection and movement tracking in video feeds.
This is the largest domain. For prebuilt Azure AI Language features, know the exact capabilities: key phrase extraction, entity recognition (including PII detection with redaction), sentiment analysis (document and sentence level with confidence scores), and language detection. Custom text classification covers single-label vs multi-label classification, data labeling in Language Studio, and model evaluation. Custom NER follows a similar pattern for domain-specific entity extraction. Conversational Language Understanding (CLU) is the successor to LUIS — know how to define intents, label entities, provide utterance examples, train, evaluate, and deploy models. Custom Question Answering (the replacement for QnA Maker) covers knowledge base creation from FAQs, manuals, and support docs, multi-turn conversations, active learning, alternate phrasing, and chit-chat. The Translator service includes text translation, transliteration, document translation, and custom translation models.
Azure AI Search is the core service here. Understand index schema design: which field attributes to use (searchable, filterable, sortable, facetable, retrievable) and when. Know the difference between simple query syntax and full Lucene query syntax. Data sources can be Azure SQL, Cosmos DB, Blob Storage, and Table Storage. Indexers automate data ingestion from these sources with scheduling options. Skillsets are the AI enrichment layer — built-in skills include OCR, image analysis, entity recognition, key phrase extraction, language detection, sentiment, translation, merge, split, and text translation. Custom skills are implemented as Azure Functions with a Web API interface. The Knowledge Store projects enriched documents to Azure Storage tables, objects, and files for downstream analytics. Know how to chain skills, use incremental enrichment, and debug skillset execution errors.
The Bot Framework SDK provides the foundation: activities (message, conversation update, event), turns, adapter (CloudAdapter), and middleware pipeline. Know how to create an EchoBot and understand the activity handler methods (OnMessageActivityAsync, OnMembersAddedAsync). Waterfall dialogs and component dialogs are key patterns for multi-turn conversations. Adaptive Cards support rich UI with Hero, Thumbnail, Fact, and Input cards with submit actions. Integrate Azure AI services by calling Language, Vision, or Speech APIs from bot middleware or dialogs. Application Insights integration provides telemetry for monitoring bot usage and errors. Bot Framework Composer offers a visual alternative for building bots with adaptive dialogs (event-driven triggers and actions), Language Generation (LG templates for responses), and publishing to Azure. Channel-specific considerations include Teams (adaptive cards fully supported, messaging extensions), Web Chat (Direct Line token authentication), and third-party channels. Bot Emulator is used for local testing.
Azure OpenAI Service provides access to GPT-4o, GPT-4 Turbo, GPT-3.5-Turbo, Embeddings (text-embedding-ada-002, text-embedding-3-small/large), and DALL-E for image generation. Know how to provision a resource, deploy models (Global Standard, Standard, Provisioned Managed), and manage quota. Prompt engineering fundamentals: system message (assistant persona), user message (input), assistant message (response). Temperature controls randomness (0 = deterministic, 1 = creative), top_p nucleus sampling, max_tokens limits response length, and frequency/presence penalties reduce repetition. Content filtering covers hate, sexual, violence, self-harm categories with severity levels (safe, low, medium, high). RAG (Retrieval Augmented Generation) combines AI Search vector search with GPT to provide grounded, factual answers based on your data. Function calling enables GPT to invoke structured APIs. Fine-tuning customizes model behavior on domain-specific data. DALL-E generates images from natural language descriptions.
You are building an application that extracts text from scanned PDF invoices. The invoices contain both printed text and handwritten notes. Which Azure AI service should you use?
A. Azure AI Vision Read API
B. Azure AI Document Intelligence prebuilt invoice model
C. Azure AI Language entity recognition
D. Azure AI Search OCR skill
Answer: B. Document Intelligence prebuilt invoice model is designed for structured document extraction from invoices, handling both printed and handwritten text. The Read API extracts text but does not understand invoice structure. The OCR skill in Search is for indexing pipelines, not direct application use.
Your Azure AI Search solution needs to extract entities from documents during indexing and store the enriched results in Azure Storage for downstream Power BI reports. What should you configure?
A. Custom analyzer and scoring profile
B. Indexer with field mappings
C. Skillset with entity recognition skill and knowledge store projections
D. Synonyms map and suggesters
Answer: C. A skillset chains AI skills like entity recognition during indexing. The knowledge store projects enriched documents to Azure Storage tables/blobs for external consumption.
Your bot needs to handle interruptions — a user should be able to change their order mid-conversation without losing context. Which dialog pattern should you use?
A. Waterfall dialog with step-by-step prompts
B. Adaptive dialog with event triggers
C. Component dialog with nested dialogs
D. Prompt dialog with validators
Answer: B. Adaptive dialogs are designed for event-driven, interruptible conversations. They allow users to change topics mid-flow while preserving conversation context. Waterfall dialogs are sequential and do not handle interruptions well.
You need to deploy an Azure AI Language service container to an IoT device that is occasionally disconnected from the internet. How should you configure billing?
A. Set the BILLING environment variable to the service endpoint
B. Use a local billing file with prepaid quota
C. Set the ApiKey environment variable to the service key
D. Configure the container with a license file
Answer: A. Containers require the Billing environment variable set to the Azure resource endpoint URL. They must periodically report usage to Azure; disconnected operation is not supported — containers require at least periodic connectivity for billing.
Your application calls Azure OpenAI GPT-4 and you notice inconsistent responses. You want the model to always respond in JSON format with specific fields. Which approach should you use?
A. Set temperature to 1.0 and top_p to 0.95
B. Use system message with JSON format instruction and set response_format to json_object
C. Post-process the response with regex parsing
D. Set max_tokens to limit output length
Answer: B. The response_format parameter set to json_object instructs the model to output valid JSON. Combined with a clear system message describing the expected schema, this provides the most reliable structured output. Temperature should be set low (0) for deterministic results.
Your application needs to detect personally identifiable information (PII) such as phone numbers and email addresses in user-submitted text and redact it before storage. Which API should you call?
A. Azure AI Vision OCR
B. Azure AI Language PII detection with redaction
C. Azure AI Content Safety text moderation
D. Azure AI Translator detect
Answer: B. The Azure AI Language PII detection API can identify and optionally redact PII entities from text. The redact parameter replaces detected entities with asterisks. Content Safety detects harmful content, not PII.
Your Azure AI Search index must support real-time spelling correction for user queries. Which feature should you enable?
A. Scoring profile with freshness boosting
B. Custom analyzer with phonetic tokenizer
C. Suggester with fuzzy search
D. Semantic search configuration
Answer: C. A suggester enables autocomplete and search-as-you-type suggestions. Combining it with fuzzy search (~ operator) provides spelling correction by matching terms within a small edit distance. Semantic search handles concept matching, not spelling.
You need to identify the language of customer support tickets before routing them to the appropriate team. Which Azure AI Language feature should you use?
A. Entity recognition
B. Language detection
C. Key phrase extraction
D. Sentiment analysis
Answer: B. Language detection identifies the primary language of input text and returns a language code and confidence score. The other options operate on different aspects: entity recognition finds named entities, key phrase extraction identifies main talking points, and sentiment analysis determines positive/negative tone.
Your Custom Vision object detection model shows high precision but low recall. What does this indicate?
A. The model is too conservative — when it predicts, it is usually correct, but it misses many objects
B. The model predicts too many false positives
C. The training dataset has too many images
D. The evaluation metric is calculated incorrectly
Answer: A. High precision means most predictions are correct (few false positives). Low recall means the model misses many actual objects (many false negatives). To improve recall, add more training data with the missed object variations, adjust the probability threshold, or retrain with more diverse examples.
You are building a RAG solution with Azure AI Search and Azure OpenAI. You want the model to answer only from your indexed documents and refuse to answer questions outside that scope. Which configuration is essential?
A. Set temperature to 0 and include a system message restricting knowledge scope
B. Use GPT-4 Turbo with higher max_tokens
C. Enable semantic search in Azure AI Search
D. Configure content filtering with strict severity levels
Answer: A. A well-crafted system message instructing the model to answer only from provided context, combined with temperature set to 0 for deterministic behavior, is the foundation of grounded RAG. The system message should say "Answer only from the provided context. If the answer is not in the context, say you don't know."
| Domain | Recommended Lab |
|---|---|
| Plan and Manage | Create an Azure AI resource and configure diagnostic logging |
| Vision | Analyze images with Computer Vision and train a Custom Vision classifier |
| NLP | Extract entities and sentiment with Language service; build a CLU app |
| Knowledge Mining | Create an AI Search index with built-in skills enrichment |
| Conversational AI | Build an EchoBot and publish to Azure; design a bot in Composer |
| Generative AI | Deploy GPT-4 and build a RAG solution with AI Search |