AI-102 Advanced Quiz - Free Practice Quiz

Question 1 of 60

An Azure AI Engineer is building a high-availability RAG solution for a global enterprise. Which architecture MOST effectively ensures sub-200ms query latency for users across three continents?

A Deploying Azure AI Search and Azure OpenAI in multiple regions with Azure Front Door for global load balancing and latency-based routing, with replicated Azure AI Search indexes per region B Single region Azure AI Search and Azure OpenAI with CDN caching of static content C One global Azure AI Search index with Azure Traffic Manager routing D Using Azure Cosmos DB as the vector store instead of Azure AI Search

1 / 60

Question 2 of 60

On the Azure AI Engineer Associate AI-102 exam, a solution architect must enable Azure OpenAI to access a private, on-premises knowledge base without exposing data to the public internet. Which architecture MOST securely achieves this?

A Storing knowledge base documents in a public Azure Blob container B Accessing the knowledge base via public internet with HTTPS encryption C Using Azure DNS private zones without VNet integration D Private Endpoint for Azure OpenAI + Azure VNet integration connecting to on-premises via ExpressRoute or VPN â€” ensuring all data stays on private network

2 / 60

Question 3 of 60

On the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively maintains conversation context across multiple user sessions in a stateless Azure OpenAI API environment?

A Persisting conversation turns in Azure Cosmos DB with a session ID key, retrieving and prepending to the messages array on each API call â€” with summarization of older turns when approaching the context limit B Using browser cookies to store all conversation history C Storing conversation history in the session token D Relying on the model's internal memory

3 / 60

Question 4 of 60

On the Azure AI Engineer Associate AI-102 exam, which advanced approach MOST effectively validates that an Azure AI Search indexer's cognitive skill pipeline produces expected enrichment output?

A Using the Debug Sessions feature in Azure AI Search â€” testing the enrichment pipeline against a sample document and inspecting per-skill input/output to identify enrichment failures B Reviewing the index schema document after indexing C Trusting cognitive skills to always produce correct output D Checking only that the indexer completes without HTTP errors

4 / 60

Question 5 of 60

According to the Azure AI Engineer Associate AI-102 exam, an enterprise requires that all Azure OpenAI prompts and completions be logged for compliance auditing. Which combination MOST comprehensively captures this data?

A Enabling Azure OpenAI diagnostic settings to route logs to Log Analytics, combined with Azure Monitor Workbooks for query and visualization â€” capturing all request/response metadata and optionally content with appropriate data protection policies B Storing logs only in application code memory C Using Azure Key Vault to log all API keys D Reviewing Azure Portal usage statistics monthly

5 / 60

Question 6 of 60

The Azure AI Engineer Associate AI-102 exam covers the 'Semantic Kernel' framework. Which capability does it provide for Azure AI engineers?

A An orchestration framework for building AI agents that chain Azure OpenAI calls, native functions, and plugins with memory and planning capabilities B A framework for training custom neural networks from scratch C A monitoring tool for Azure AI Services usage D A data labeling interface for custom vision models

6 / 60

Question 7 of 60

The Azure AI Engineer Associate AI-102 exam covers which strategy MOST effectively manages the freshness of a RAG knowledge base when source documents are updated frequently?

A Implementing incremental indexing â€” using Azure AI Search indexers with change-detection policies to reindex only updated documents, combined with cache purging for stale cached responses B Using only static documents that never change C Relying on users to clear their browser cache D Rebuilding the full Azure AI Search index from scratch daily

7 / 60

Question 8 of 60

The Azure AI Engineer Associate AI-102 exam covers which advanced Azure AI Content Safety feature allows defining organizational-specific prohibited topics for Azure OpenAI?

A Groundedness detection combined with protected material detection and custom category classifiers trained on domain-specific prohibited content examples B Deploying a separate moderation API endpoint C Using Azure Policy to restrict Azure OpenAI usage D Increasing Azure OpenAI's built-in filter sensitivity only

8 / 60

Question 9 of 60

On the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively implements responsible AI content filtering for a customer-facing Azure OpenAI application that must comply with EU AI Act risk requirements?

A Combining Azure AI Content Safety filters (configured for the application's risk level), Prompt Shield for jailbreak detection, custom blocklists for domain-specific prohibited content, and human-in-the-loop review for high-risk outputs â€” with documented risk assessment B Using only the default Azure OpenAI content filter without customization C Relying on users to report inappropriate content after the fact D Blocking all user input longer than 100 tokens

9 / 60

Question 10 of 60

According to the Azure AI Engineer Associate AI-102 exam, which advanced Azure AI Search optimization MOST improves recall for technical documentation queries where synonyms and abbreviations are common?

A Increasing the index size by adding redundant documents B Adding more vector dimensions to the embedding model C Enabling faceted navigation for technical categories D Configuring synonym maps (e.g., 'AI' â†” 'Artificial Intelligence') and custom analyzers for technical terms â€” combined with hybrid search for broad recall

10 / 60

Question 11 of 60

According to the Azure AI Engineer Associate AI-102 exam, which Azure AI advanced feature enables creating a custom voice that mimics a specific person's speech patterns for TTS applications?

A Custom Neural Voice â€” training a TTS model on a person's recorded speech samples with written consent, creating a unique branded voice B Speaker recognition enrollment C Increasing speech synthesis speed D Standard neural voice with different voice options

11 / 60

Question 12 of 60

According to the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively implements multi-agent orchestration using Azure OpenAI, where specialized agents collaborate to complete a complex task?

A Using Semantic Kernel's Agent framework or AutoGen â€” orchestrating multiple specialized Azure OpenAI agents (e.g., planner, researcher, writer, critic) that communicate and refine outputs collaboratively B Calling Azure OpenAI sequentially with the same system prompt for each subtask C Deploying one agent for all users without specialization D Using a single monolithic prompt for all tasks

12 / 60

Question 13 of 60

The Azure AI Engineer Associate AI-102 exam covers fine-tuning Azure OpenAI models. Which scenario MOST justifies the additional cost and complexity of fine-tuning over prompt engineering?

A When the model needs to consistently produce a very specific output format or domain-specific style that cannot be reliably achieved through few-shot prompting, and where the latency savings from shorter prompts justify the fine-tuning investment B When a one-time Q&A task needs to be completed for a small dataset C When the application only needs to summarize documents once per week D When adding a new language the model has never seen before

13 / 60

Question 14 of 60

On the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively handles a scenario where Azure AI Document Intelligence produces low-confidence extractions on degraded-quality scanned documents?

A Rejecting all documents with any low-confidence field B Accepting all low-confidence extractions without review C Using only the Read model OCR output without structured extraction D Implementing confidence-threshold-based human review routing â€” low-confidence field extractions automatically routed to a human review queue via Azure Logic Apps or Power Automate

14 / 60

Question 15 of 60

On the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively handles a high-volume Azure AI Search application experiencing throttling (HTTP 429) errors during peak traffic?

A Implementing exponential backoff retry logic, scaling the search service replicas to increase throughput capacity, and using caching to reduce repeat queries B Switching to a Free tier with more throughput C Reducing search query complexity to zero results D Retrying immediately without backoff

15 / 60

Question 16 of 60

On the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively handles Azure OpenAI rate limit errors (429) in a production application serving thousands of concurrent users?

A Failing immediately on every 429 error B Implementing retry with exponential backoff and jitter, combining with a queue-based workload leveling pattern (Azure Service Bus) and distributing load across multiple Azure OpenAI deployments in different regions C Increasing API key count without quota increase D Alerting developers manually when 429s appear

16 / 60

Question 17 of 60

According to the Azure AI Engineer Associate AI-102 exam, which advanced Azure AI Search feature enables searching for products by uploading an image rather than a text query?

A Full-text search with image metadata keywords B Multimodal vector search using Azure AI Vision image embeddings stored as vectors in Azure AI Search â€” querying with an image embedding against indexed product image vectors C OCR-based text extraction from product images for text search D Azure Custom Vision product classification as a search pre-filter

17 / 60

Question 18 of 60

The Azure AI Engineer Associate AI-102 exam identifies which approach as MOST appropriate for building an AI solution that autonomously completes multi-step tasks using Azure OpenAI?

A Single-turn API calls with no memory or state B Pre-scripted rule-based decision trees C Traditional chatbot with fixed responses D Agentic AI pattern using Azure OpenAI function calling with tool use â€” the model decides which tools to invoke, in what order, iterating until the task is complete

18 / 60

Question 19 of 60

The Azure AI Engineer Associate AI-102 exam identifies which advanced Azure AI Document Intelligence scenario as requiring the table extraction capability?

A Recognizing handwritten signatures on documents B Extracting financial data from earnings reports where values are organized in multi-row, multi-column table structures requiring row/column relationship preservation C Identifying document type for classification routing D Detecting document language for translation routing

19 / 60

Question 20 of 60

The Azure AI Engineer Associate AI-102 exam identifies which advanced Azure AI Document Intelligence capability enables training a model on only a few (5â€“50) labeled samples while achieving high accuracy?

A Zero-shot extraction using the Read model B Custom neural extraction model with template and neural training â€” the AI Document Intelligence custom model learns structural patterns efficiently from small labeled datasets C Rule-based regex extraction from OCR output D General Document model with default extraction

20 / 60

Question 21 of 60

The Azure AI Engineer Associate AI-102 exam identifies which advanced pattern as MOST effectively preventing prompt injection attacks in an Azure OpenAI application that processes user-provided documents?

A Trusting all user document content as safe by default B Using Azure AI Content Safety Prompt Shield to detect injection attempts, sandboxing document content from system instructions using clear delimiters, and validating model output against expected schemas before use C Only allowing users to upload PDFs smaller than 1 MB D Blocking documents with more than 1,000 words

21 / 60

Question 22 of 60

According to the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively manages Azure OpenAI usage limits across multiple application teams sharing a single Azure OpenAI resource?

A Giving all teams the same API key without rate limiting B Blocking all teams except the highest-priority one C Manually tracking token usage in a spreadsheet D Azure API Management with per-team subscription keys, rate limiting policies, and quota enforcement â€” routing to the shared Azure OpenAI resource with fair allocation

22 / 60

Question 23 of 60

According to the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively enables an Azure AI Language custom NER model to improve performance on underrepresented entity types?

A Deleting all underperforming entity types from the schema B Targeted data collection and annotation of additional utterances containing the underperforming entity type â€” guided by the model's per-entity evaluation metrics C Reducing the model's confidence threshold to accept more predictions D Retraining from scratch with a completely new dataset

23 / 60

Question 24 of 60

According to the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively evaluates an Azure Bot Service bot's end-to-end conversational quality before production deployment?

A Testing only the bot's connection to the Teams channel B End-to-end conversation testing using the Bot Framework Emulator with scripted dialogue scenarios, combined with automated transcript evaluation using LLM-as-judge scoring of response appropriateness, completeness, and tone C Checking the bot's HTTP endpoint health only D Reviewing LUIS training data only

24 / 60

Question 25 of 60

On the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively evaluates an Azure OpenAI RAG system's response quality at scale without human annotation for every response?

A Manual human review of all 10,000 daily responses B LLM-as-judge evaluation â€” using a separate, more capable Azure OpenAI model to score responses on dimensions like groundedness, relevance, coherence, and fluency against the retrieved context C Using only ROUGE score for all evaluation scenarios D Measuring only the Azure AI Search retrieval latency

25 / 60

Question 26 of 60

The Azure AI Engineer Associate AI-102 exam covers which AI-102 advanced scenario as MOST suitable for Azure AI Vision's 'product recognition' capability in retail?

A Generating written product descriptions from images B Training a custom speech model for retail use C Detecting shop inventory levels by counting visible shelf gaps D Identifying specific products on retail shelves by visual similarity using model customization â€” enabling automated inventory and planogram compliance

26 / 60

Question 27 of 60

On the Azure AI Engineer Associate AI-102 exam, which advanced Azure OpenAI pattern enables an application to process user requests in a streaming fashion â€” displaying partial responses as they are generated?

A Setting max_tokens to a very small value B Using the Azure OpenAI streaming API (stream: true) â€” consuming server-sent events and progressively rendering tokens as they arrive for a responsive user experience C Setting temperature to 0 for deterministic output D Polling the API every 100ms for new content

27 / 60

Question 28 of 60

On the Azure AI Engineer Associate AI-102 exam, which Azure AI advanced pattern enables a generative AI application to cite sources for every factual claim in its response?

A Fine-tuning the model to always add citations B Inline citation prompting with grounding enforcement â€” instructing the model in the system prompt to cite specific retrieved document excerpts for each claim, then validating citations against the actual source chunks C Asking users to verify facts themselves D Setting a high presence penalty to reduce repetition

28 / 60

Question 29 of 60

According to the Azure AI Engineer Associate AI-102 exam, which Azure AI Document Intelligence advanced feature allows building a composed model that routes documents to specialized sub-models based on document type?

A Using a single general document model for all types B Composed models â€” combining multiple custom models so the API automatically selects the appropriate sub-model based on document type classification C Building separate API endpoints per document type D Manual routing based on file extension

29 / 60

Question 30 of 60

On the Azure AI Engineer Associate AI-102 exam, which technique MOST effectively chunks documents for Azure AI Search RAG indexing to balance retrieval precision and context richness?

A Fixed-character chunking (every 100 characters regardless of content) B Splitting documents only at file page boundaries C Semantic chunking â€” splitting at natural semantic boundaries (paragraphs, sections) and overlapping adjacent chunks to preserve cross-boundary context D Indexing each document as a single chunk regardless of length

30 / 60

Question 31 of 60

The Azure AI Engineer Associate AI-102 exam covers which approach MOST effectively isolates a test environment Azure OpenAI deployment from a production deployment?

A Using different temperature settings for test vs. production B Sharing one deployment between test and production with different API keys C Creating separate Azure OpenAI resources in separate resource groups per environment (dev, staging, prod) â€” each with independent deployments, quotas, and access controls D Deploying to the same resource with different model versions

31 / 60

Question 32 of 60

The Azure AI Engineer Associate AI-102 exam covers which approach MOST effectively enables Azure AI Vision to detect defects in manufacturing parts with very limited labeled training images?

A Using Azure AI Vision's pre-built object detection without custom training B Running the Read OCR API on part images C Few-shot custom model training using Azure AI Custom Vision with the General (compact) domain and aggressive data augmentation (horizontal/vertical flips, brightness and contrast variation, rotation) to maximize model generalization from limited labeled images D Using the standard Image Analysis captioning feature

32 / 60

Question 33 of 60

The Azure AI Engineer Associate AI-102 exam covers which approach MOST effectively minimizes token costs while maintaining response quality for a mixed-complexity Azure OpenAI workload?

A Using the smallest model for all queries regardless of complexity B Using the largest model for all queries regardless of complexity C Implementing prompt routing â€” classifying user queries by complexity and routing simple queries to a smaller, cheaper model (e.g., GPT-3.5-Turbo) and complex queries to a larger model (e.g., GPT-4) D Sending all queries as batch requests regardless of user-facing latency requirements

33 / 60

Question 34 of 60

According to the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively tests that an Azure Bot Service bot handles unexpected user inputs gracefully?

A Only testing the happy path with expected inputs B Reviewing bot dialog code manually without execution C Testing only the bot's integration with Azure AI Services APIs D Automated load and boundary testing using Bot Framework Emulator and automated testing frameworks, with edge case inputs (empty strings, very long inputs, off-topic queries)

34 / 60

Question 35 of 60

According to the Azure AI Engineer Associate AI-102 exam, which advanced Azure AI Content Safety feature enables detecting specific types of violent extremist content in user-generated text?

A Standard hate speech detection only B Azure Monitor log analysis C Custom categories in Azure AI Content Safety â€” training a classifier on domain-specific harmful content types beyond the four default harm categories D Language detection

35 / 60

Question 36 of 60

According to the Azure AI Engineer Associate AI-102 exam, which advanced approach provides the MOST comprehensive cost governance for an enterprise Azure OpenAI deployment serving multiple business units?

A Using a single shared Azure OpenAI resource with no cost tracking B Restricting all business units to one model deployment C Tagging Azure OpenAI resources by business unit, implementing per-unit Provisioned Throughput Unit (PTU) reservations, using Azure Cost Management with tag-based budget alerts, and surfacing token consumption dashboards via Azure Monitor Workbooks D Providing each business unit with direct Azure OpenAI access without controls

36 / 60

Question 37 of 60

On the Azure AI Engineer Associate AI-102 exam, which pattern MOST effectively handles the Azure OpenAI context window limit when processing very long documents for summarization?

A Sending the full document in a single API call regardless of token count B Truncating the document to fit the context window C Map-reduce summarization â€” splitting the document into chunks that fit within the context window, summarizing each chunk independently, then synthesizing the chunk summaries into a final summary D Using Azure AI Search to index the document and skipping summarization

37 / 60

Question 38 of 60

The Azure AI Engineer Associate AI-102 exam identifies which approach as enabling an Azure OpenAI model to use real-time external data that postdates its training cutoff?

A Fine-tuning the model with recent data each week B Reducing the context window size to force recency C Increasing the model's temperature to generate more up-to-date guesses D Augmenting the model's context with dynamically retrieved real-time data using tools (function calling to external APIs or Azure AI Search) â€” enabling knowledge beyond training cutoff

38 / 60

Question 39 of 60

On the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively enables evaluating the factual accuracy of Azure OpenAI RAG responses using automated metrics?

A Measuring only response length and latency B Asking users to rate each response manually C Groundedness evaluation â€” measuring whether each claim in the model's response is supported by the retrieved source documents using an LLM-as-judge or NLI model D Using perplexity scores as the sole accuracy metric

39 / 60

Question 40 of 60

On the Azure AI Engineer Associate AI-102 exam, which advanced Azure AI Language technique MOST effectively improves sentiment analysis accuracy for domain-specific text that contains industry jargon?

A Using only the pre-built general sentiment model B Reducing the confidence threshold for all predictions C Fine-tuning a custom sentiment model using Azure AI Language's custom text classification with domain-specific labeled examples that include industry jargon correctly annotated for sentiment D Relying on key phrase extraction as a proxy for sentiment

40 / 60

Question 41 of 60

According to the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively validates that an Azure AI Language CLU model meets production quality requirements before deployment?

A Measuring only overall accuracy on the training set B Deploying to production immediately after training C Evaluating the model against a held-out test set with precision, recall, F1 disaggregated by intent and entity type â€” establishing minimum thresholds for each, and using the AI-102 recommended iterative improvement loop (add utterances â†’ retrain â†’ evaluate) D Testing only the model's response time

41 / 60

Question 42 of 60

On the Azure AI Engineer Associate AI-102 exam, which Azure AI Language advanced feature is MOST appropriate for analyzing legal contracts to identify all parties, obligations, and key dates?

A Pre-built NER with standard entity types only B Standard sentiment analysis on contract text C Custom NER trained on labeled legal contracts â€” identifying domain-specific entities like parties, obligations, governing law, and contract dates D Key phrase extraction from legal documents

42 / 60

Question 43 of 60

The Azure AI Engineer Associate AI-102 exam identifies which Azure security feature prevents Azure AI Services endpoints from being accessed from unauthorized networks while still allowing Azure service access?

A Azure AD Conditional Access for users only B Azure Policy for resource tags C Azure Private Endpoint combined with network access rules that disable public access â€” routing Azure AI Services traffic exclusively through private VNet endpoints D Azure Role-Based Access Control (RBAC) alone

43 / 60

Question 44 of 60

The Azure AI Engineer Associate AI-102 exam identifies which advanced approach as enabling Azure AI Search to serve personalized search results based on an individual user's interaction history?

A Adding a user's name as a filter in every query B Using only global popularity ranking C Incorporating user interaction signals (clicks, dwell time) as feedback to re-rank results â€” using Azure AI Search scoring profiles or a learned ranking model trained on interaction data D Sorting results by document creation date

44 / 60

Question 45 of 60

The Azure AI Engineer Associate AI-102 exam covers which Azure AI approach MOST effectively enables a bot to escalate to a live human agent when it cannot confidently handle a user's request?

A Providing the human agent's email address for the user to contact separately B Ending the conversation immediately when confidence is low C Implementing handoff to human agents using the Bot Framework's handoff protocol â€” triggering when intent confidence falls below a threshold, preserving full conversation context for the receiving agent D Restarting the conversation from the beginning

45 / 60

Question 46 of 60

According to the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively reduces Azure AI Search index storage costs for a large corpus with many rarely queried documents?

A Storing all documents in the highest-cost storage tier B Using tiered indexing â€” keeping frequently accessed documents in a standard search index while archiving rarely accessed documents to a Storage Optimized index or separate cold storage C Disabling semantic search to reduce processing costs D Increasing the number of replicas for the full index

46 / 60

Question 47 of 60

According to the Azure AI Engineer Associate AI-102 exam, which Bot Framework advanced pattern enables a bot to invoke external REST APIs and use their results within a conversation?

A Hard-coding API responses in dialog LUIS B Using Bot Framework Emulator to simulate API calls C Configuring static QnA answers that include API URLs D Implementing service clients within Bot Framework Dialogs â€” calling external REST APIs from dialog step methods and incorporating API responses into the bot's reply

47 / 60

Question 48 of 60

According to the Azure AI Engineer Associate AI-102 exam, which advanced strategy MOST effectively validates that an Azure AI Document Intelligence custom model generalizes to document samples outside the training distribution?

A Evaluating on the same samples used for training B Accepting the model immediately after training completes C Using the model confidence score alone as a validation metric D Held-out test set evaluation with a diverse sample representing edge cases (poor scan quality, unusual layouts, multilingual content) â€” combined with human review of borderline predictions

48 / 60

Question 49 of 60

On the Azure AI Engineer Associate AI-102 exam, which deployment configuration ensures that Azure AI Search does NOT experience downtime during an index schema update?

A Using the Free tier which allows instant schema changes B Deleting and recreating the index during maintenance windows C Updating the index schema in place during peak traffic D Using index aliases to maintain two indexes (current and updated) and atomically switching the alias to the updated index after successful backfill

49 / 60

Question 50 of 60

The Azure AI Engineer Associate AI-102 exam covers which advanced Bot Framework pattern enables a bot to proactively send messages to users without them initiating the conversation?

A Using turn context from an incoming message to send outbound messages B Proactive messaging â€” using a stored Conversation Reference to initiate a conversation from outside a message handler, triggered by external events C Relying only on channel-initiated conversations D Configuring a webhook that only responds to user messages

50 / 60

Question 51 of 60

On the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively reduces hallucination risk when using Azure OpenAI for medical question answering?

A Using the highest temperature setting for creative medical answers B Relying on the model's pre-training medical knowledge alone C Setting frequency penalty to maximum D RAG with verified medical knowledge bases combined with strict system prompts instructing the model to only answer from retrieved context and explicitly acknowledge when the answer is unknown

51 / 60

Question 52 of 60

On the Azure AI Engineer Associate AI-102 exam, which design pattern MOST effectively prevents an Azure OpenAI system from producing harmful outputs in response to indirect prompt injections embedded in retrieved RAG documents?

A Only using documents from trusted internal sources B Disabling the model's ability to use retrieved documents C Setting temperature to 0 to restrict output variability D Sanitizing retrieved document content before injection into the prompt, using Azure AI Content Safety Prompt Shield to detect injection payloads, and validating model outputs against expected response schemas

52 / 60

Question 53 of 60

According to the Azure AI Engineer Associate AI-102 exam, which advanced technique improves Azure OpenAI function calling reliability for a complex multi-step agentic workflow?

A Setting temperature to 1.0 for maximum creativity B Calling all functions simultaneously in a single prompt C Using free-form text outputs instead of function calling for flexibility D Structured outputs (JSON mode) combined with explicit step-by-step function call planning in the system prompt, validation of each function call result before proceeding, and retry logic for malformed outputs

53 / 60

Question 54 of 60

On the Azure AI Engineer Associate AI-102 exam, which Azure AI approach MOST effectively detects when an Azure AI Language CLU model has been exposed to adversarial inputs designed to change its intent classification?

A Reviewing model accuracy on the training set B Implementing confidence score thresholding â€” flagging low-confidence intent classifications for review and logging unusual input patterns for adversarial input analysis C Trusting all CLU outputs without validation D Checking API response times for spikes

54 / 60

Question 55 of 60

The Azure AI Engineer Associate AI-102 exam covers which advanced Azure AI Vision scenario as MOST appropriate for detecting objects in real-time video with bounding box tracking across frames?

A Azure AI Document Intelligence's OCR on video frames B Azure Custom Vision batch image classification C Azure Video Indexer summarization of recorded video D Custom object detection model with Azure AI Vision spatial analysis or Azure Custom Vision real-time inference integrated with video frame extraction pipeline

55 / 60

Question 56 of 60

The Azure AI Engineer Associate AI-102 exam covers which advanced Azure AI capability enables analyzing sentiment and topics across thousands of customer survey responses at scale with zero labeled data?

A Custom NER trained on survey text only B Standard key phrase extraction only C Azure OpenAI GPT-4 with temperature 0 applied to each survey response individually D Summarization and clustering of survey responses using Azure OpenAI embeddings + Azure AI Search vector clustering to identify emergent themes, combined with sentiment analysis from Azure AI Language

56 / 60

Question 57 of 60

The Azure AI Engineer Associate AI-102 exam identifies which approach as MOST effectively monitoring a production RAG application for retrieval failures (cases where no relevant documents are found)?

A Using only user satisfaction surveys as the quality signal B Monitoring only API error rates (HTTP 5xx) C Reviewing application logs manually each week D Tracking the percentage of responses where Azure AI Search returns zero results or where the Azure OpenAI model indicates it cannot find relevant information in the context â€” alerting when the no-answer rate exceeds a defined SLO threshold

57 / 60

Question 58 of 60

According to the Azure AI Engineer Associate AI-102 exam, which advanced pattern MOST effectively enables an Azure AI engineer to roll back a deployed Azure OpenAI model version without application downtime?

A Deleting the current deployment and creating a new one B Maintaining multiple Azure OpenAI deployments (current and previous versions) and using Azure API Management or application-level routing to switch traffic between deployments without downtime C Hardcoding model versions in application code D Always deploying new model versions to the same deployment name

58 / 60

Question 59 of 60

According to the Azure AI Engineer Associate AI-102 exam, which advanced pattern MOST effectively manages prompt template versioning in a production Azure OpenAI application?

A Hard-coding prompt templates in application code B Storing prompts only in developer notebooks C Sharing prompts via email between team members D Storing prompt templates in a version-controlled repository (e.g., Azure DevOps), deploying via CI/CD, and including prompt version metadata in every Azure OpenAI API call log â€” enabling prompt change attribution for quality regression investigation

59 / 60

Question 60 of 60

According to the Azure AI Engineer Associate AI-102 exam, which approach MOST effectively implements conversation-level abuse detection for an Azure OpenAI chatbot serving a public audience?

A Reviewing chat logs only after a user complaint is received B Blocking all conversations exceeding 1,000 tokens C Using character-level keyword blocking only D Per-turn Azure AI Content Safety filtering combined with session-level anomaly detection that flags accounts exhibiting patterns of repeated policy violations â€” triggering automated account-level review

60 / 60