Your organization is building a knowledge base from technical manuals stored in multiple formats (PDF, Word, HTML). You need to extract text and images from these documents and create a searchable index. The solution must handle tables and preserve their structure. Which approach should you use?
A.Upload documents directly to Azure AI Search
B.Use Azure AI Language custom entity extraction
C.Use Azure AI Document Intelligence layout model as a custom skill
D.Use Azure AI Vision OCR skill in the skillset
AnswerC
The layout model extracts text, tables, and structure preserving relationships.
Why this answer
Azure AI Document Intelligence Layout model extracts text, tables, and their structure from documents, making it suitable for preserving table structure and building a searchable knowledge base. It can be used as a custom skill in an Azure AI Search enrichment pipeline. Option A is incorrect because uploading documents directly to Azure AI Search only indexes raw text without extraction or table structure preservation.
Option B is incorrect because Custom Entity Extraction in Azure AI Language focuses on identifying entities, not extracting tables or preserving structure. Option D is incorrect because Azure AI Vision OCR extracts text but not table structure, and is not designed for document layout analysis.
You are designing a solution that uses Azure AI Document Intelligence to extract data from invoices. The solution must classify invoices by vendor and extract line items. Which prebuilt model should you use?
A.Prebuilt invoice model
B.Custom extraction model
C.Prebuilt receipt model
D.Prebuilt layout model
AnswerA
Invoice model extracts vendor and line items.
Why this answer
The Prebuilt invoice model (Option A) is specifically designed to extract common fields from invoices, including vendor details and line items, without requiring custom training. This model is optimized for invoice documents and provides out-of-the-box extraction of structured data such as vendor name, invoice date, and line-item descriptions, quantities, and amounts.
Exam trap
The trap here is that candidates often confuse the Prebuilt layout model with the Prebuilt invoice model, assuming layout extraction is sufficient for invoice data, but layout only provides raw text and table positions without the semantic understanding needed for vendor classification and line-item extraction.
How to eliminate wrong answers
Option B is wrong because a custom extraction model requires labeled training data and is used when prebuilt models do not meet specific document needs, but here the prebuilt invoice model already covers vendor classification and line-item extraction. Option C is wrong because the Prebuilt receipt model is designed for receipts (e.g., from stores or restaurants), not invoices, and does not extract vendor classification or line-item details in the same structured format. Option D is wrong because the Prebuilt layout model extracts text, tables, and selection marks but does not perform semantic classification or vendor-specific field extraction; it lacks the prebuilt understanding of invoice-specific fields.
Refer to the exhibit. You are configuring an Azure OpenAI Service deployment for document summarization. The current parameters produce summaries that are often too verbose. You need to make the summaries more concise while maintaining factual accuracy. Which parameter change should you make?
A.Increase top_p to 1.0
B.Increase frequency_penalty to 0.5
C.Decrease max_tokens to 100
D.Increase temperature to 0.7
AnswerC
Reducing max_tokens limits output length, making summaries more concise.
Why this answer
Decreasing max_tokens to 100 directly limits the maximum length of the generated summary, forcing the model to produce shorter output. This addresses the verbosity issue without altering the model's factual accuracy, as max_tokens controls output length, not content selection or creativity.
Exam trap
Microsoft often tests the distinction between parameters that control output length (max_tokens) versus those that control creativity or diversity (temperature, top_p, frequency_penalty), leading candidates to mistakenly adjust the latter when the issue is simply excessive length.
How to eliminate wrong answers
Option A is wrong because increasing top_p to 1.0 makes the model consider a wider set of possible tokens, which can increase diversity and potentially lead to even more verbose or less focused summaries. Option B is wrong because increasing frequency_penalty to 0.5 penalizes tokens that have already appeared, reducing repetition but not directly controlling summary length; it may even cause the model to use more unique words, increasing verbosity. Option D is wrong because increasing temperature to 0.7 increases randomness in token selection, which can produce more creative but less consistent summaries, potentially harming factual accuracy and not reliably reducing length.
You are deploying a chatbot using Azure OpenAI Service with a custom dataset indexed in Azure AI Search. Users report that the chatbot frequently responds with 'I don't know' for questions that the dataset should cover. What is the most likely cause?
A.The search scope is limited to a small number of documents.
B.The confidence threshold in the retrieval configuration is set too low, filtering out relevant chunks.
C.The temperature setting in the model deployment is set too high.
D.The chunk size in the index is too large, causing irrelevant chunks to be retrieved.
AnswerB
Low confidence threshold causes relevant chunks to be excluded, leading to 'I don't know' responses.
Why this answer
When the confidence threshold is set too low, Azure AI Search filters out retrieved chunks that don't meet the minimum confidence score, even if those chunks are relevant. This causes the chatbot to respond with 'I don't know' because no sufficiently confident context is passed to the Azure OpenAI model for answer generation. The issue is specifically in the retrieval configuration, not in the model's generation parameters.
Exam trap
The trap here is that candidates often confuse the confidence threshold in retrieval with the temperature parameter in generation, assuming a high temperature causes the model to refuse answers, when in fact temperature controls creativity, not retrieval filtering.
How to eliminate wrong answers
Option A is wrong because limiting the search scope to a small number of documents would reduce the pool of potential matches, but the chatbot would still return answers from those documents if they contain relevant information; the symptom of 'I don't know' for covered questions points to a filtering issue, not a scope limitation. Option C is wrong because a high temperature setting affects the randomness and creativity of the generated response, not the model's ability to retrieve or use context; it might cause verbose or off-topic answers, but not a refusal to answer. Option D is wrong because large chunk sizes can cause retrieval of irrelevant chunks due to lower precision, but this would lead to incorrect or hallucinated answers, not the model saying 'I don't know'; the model would still attempt to answer using the retrieved context.
Refer to the exhibit. You are testing a conversational language understanding (CLU) application in Azure AI Language. You send the JSON request to the prediction endpoint and receive a 200 response but with an empty topIntent. What is the most likely cause?
A.The projectName or deploymentName is incorrect.
B.The conversationItem is missing a required 'role' field.
C.The language parameter is not supported.
D.The utterance text is too short.
AnswerB
The 'role' field indicates who is speaking and is required for correct classification.
Why this answer
B is correct because the JSON request for a conversational language understanding (CLU) application requires a 'role' field within the 'conversationItem' object when using the conversation analysis endpoint. Without this field, the service cannot properly interpret the utterance's context (e.g., user vs. system role), leading to a successful HTTP 200 response but an empty 'topIntent' as the model fails to classify the intent.
Exam trap
The trap here is that candidates assume a 200 status code means the request was fully valid, overlooking that CLU can return a successful HTTP response with an empty topIntent when required schema fields like 'role' are missing, rather than throwing an explicit error.
How to eliminate wrong answers
Option A is wrong because an incorrect projectName or deploymentName would result in a 404 (Not Found) or 401 (Unauthorized) error, not a 200 response with an empty topIntent. Option C is wrong because an unsupported language parameter would cause a 400 (Bad Request) error with a specific error message, not a successful 200 response. Option D is wrong because the utterance text being too short does not prevent intent classification; CLU can handle short utterances, and the model would still return a topIntent (even if low confidence) rather than an empty value.
You are debugging a CLU application. The JSON above shows a request to the Azure AI Language runtime API. The response returns an intent of "BookFlight" with a confidence of 0.95, but no entities are extracted. The training data includes entities like "Location" and "DateTime". What is the most likely cause?
A.The deployment name is incorrect.
B.The model was not trained on entity extraction.
C.The utterance does not contain any entities.
D.The request is missing the 'verbose' parameter set to true.
AnswerD
Verbose parameter is required to get entity details.
Why this answer
The CLU runtime API, by default, returns only the top intent and does not include extracted entities unless the 'verbose' query parameter is set to true. Without this parameter, the response omits the entities array even if the model was trained to extract them, which explains why the intent is returned with high confidence but no entities appear.
Exam trap
Azure often tests the misconception that a successful API response with a high-confidence intent implies all features (like entity extraction) are working correctly, when in fact the default response may omit entities unless a specific parameter is included.
How to eliminate wrong answers
Option A is wrong because an incorrect deployment name would cause a 404 or 401 error, not a successful response with a high-confidence intent. Option B is wrong because the training data includes entities like 'Location' and 'DateTime', indicating the model was trained on entity extraction; if it were not, the model would not have been configured to extract entities at all. Option C is wrong because the utterance may contain entities, but the API response does not include them due to the missing 'verbose' parameter; the absence of entities in the response does not prove they are absent from the utterance.
A healthcare organization uses Azure AI Language to analyze clinical notes. They need to detect protected health information (PHI) such as patient names and dates of birth, and also identify medical conditions. Which Azure AI Language feature should they use?
A.Conversation Analysis
B.Custom Text Classification
C.Text Analytics for health (general availability)
D.Health Text Analytics (preview) in Azure AI Language
AnswerD
This feature includes PHI detection and medical entity recognition.
Why this answer
Health Text Analytics (preview) in Azure AI Language is the correct feature because it is specifically designed to extract and structure medical information from unstructured clinical text, including protected health information (PHI) such as patient names and dates of birth, as well as medical conditions. This preview API extends the Text Analytics for health capabilities with enhanced PHI detection and entity linking, making it the appropriate choice for healthcare compliance and clinical analysis.
Exam trap
The trap here is that candidates often confuse the generally available Text Analytics for health with the preview Health Text Analytics, mistakenly thinking the GA version includes full PHI detection, when in fact the preview version is the one that explicitly adds that capability.
How to eliminate wrong answers
Option A is wrong because Conversation Analysis is designed for analyzing conversational transcripts (e.g., from chatbots or call centers) to extract intents and entities, not for detecting PHI or medical conditions in clinical notes. Option B is wrong because Custom Text Classification allows you to train a model to classify text into custom categories, but it does not natively support PHI detection or medical entity extraction out of the box. Option C is wrong because Text Analytics for health (general availability) provides medical entity extraction and relation detection but does not include the dedicated PHI detection capabilities that the preview version (Health Text Analytics) offers; the question specifically requires PHI detection, which is a key differentiator of the preview feature.
You are developing a solution to detect and redact personally identifiable information (PII) in images of documents. Which THREE Azure services can you use together?
Select 3 answers
A.Azure Custom Vision
B.Azure AI Language (PII detection)
C.Azure AI Document Intelligence
D.Azure AI Vision (Read API)
E.Azure Video Indexer
AnswersB, C, D
Detects PII in text.
Why this answer
Options B, C, and D are correct because Azure AI Document Intelligence and Azure AI Vision (Read API) can extract text from images of documents, and Azure AI Language (PII detection) can identify PII in the extracted text, enabling redaction. Option A (Custom Vision) is incorrect because it is designed for image classification and object detection, not PII detection. Option E (Video Indexer) is incorrect because it is for analyzing videos, not static images.
You need to translate a document from English to Spanish using Azure OpenAI. Which parameter should you include in the prompt to specify the target language?
A.max_tokens
B.user message
C.temperature
D.system message
AnswerB
User message should contain the translation instruction.
Why this answer
To specify the target language in an Azure OpenAI translation prompt, you include the instruction within the user message (e.g., 'Translate the following English text to Spanish: ...'). The user message is the primary input where the model receives the task and context, making it the correct parameter for language specification.
Exam trap
The trap here is that candidates often confuse the system message (which sets the assistant's role) with the user message (which provides the specific task), leading them to incorrectly choose system message for language specification.
How to eliminate wrong answers
Option A is wrong because max_tokens controls the maximum number of tokens in the response, not the target language. Option C is wrong because temperature controls the randomness of the output, not the language. Option D is wrong because the system message sets the assistant's behavior or persona (e.g., 'You are a helpful translator'), but the specific target language must be explicitly stated in the user message to direct the translation task.
A developer is using Azure Cognitive Service for Language to perform sentiment analysis on customer reviews. The service returns sentiment labels (positive, negative, neutral) and confidence scores. For a particular review, the service returns 'positive' with a confidence score of 0.55. The developer wants to ensure that only high-confidence results are used. What should the developer do?
A.Use the Text Analytics for Health API instead.
B.Retrain the sentiment analysis model with additional labeled data.
C.Configure a minimum confidence threshold of 0.75 in the application logic.
D.Adjust the input text by removing ambiguous phrases.
AnswerC
This filters out low-confidence results.
Why this answer
The developer must implement a confidence threshold in the application logic to filter out low-confidence results. The Azure Cognitive Service for Language returns confidence scores between 0 and 1 for each sentiment label, and the developer can set a minimum threshold (e.g., 0.75) to ensure only high-confidence predictions are used. This approach does not require retraining the model or modifying the input text, as the threshold is applied post-inference in the client code.
Exam trap
The trap here is that candidates may assume they can retrain the prebuilt sentiment model (Option B) or use a different API (Option A) to solve the confidence issue, when in fact the correct solution is a simple application-level threshold check.
How to eliminate wrong answers
Option A is wrong because the Text Analytics for Health API is a specialized domain-specific API for extracting medical entities and relationships, not for general sentiment analysis or confidence thresholding. Option B is wrong because the prebuilt sentiment analysis model in Azure Cognitive Service for Language cannot be retrained with custom labeled data; custom model training is only available for custom text classification or custom named entity recognition, not for the built-in sentiment analysis. Option D is wrong because removing ambiguous phrases from input text does not guarantee higher confidence scores; the model's confidence is a function of its internal weights and the entire input, and manually editing text may introduce bias or reduce the sample size without addressing the underlying confidence threshold requirement.
Refer to the exhibit. You are implementing an Azure AI Search index for semantic search with vector support. The index includes a field 'descriptionVector' of type Collection(Edm.Single) with 1536 dimensions. When you run a vector search query, you notice that results are not sorted by relevance. What is the most likely cause?
A.The 'descriptionVector' field is not searchable
B.The vector dimensions do not match the embedding model output
C.The index does not have a vector configuration with a similarity metric
D.The 'descriptionVector' field is set as retrievable: false
AnswerC
Vector search requires a vector profile to compute similarity.
Why this answer
Azure AI Search requires a vector configuration with a similarity metric (e.g., cosine, dotProduct, euclidean) to compute relevance scores for vector search results. Without this configuration, the search engine cannot sort results by relevance, leading to unsorted or default ordering. The similarity metric is defined in the index's vector configuration and is essential for ranking vector query results.
Exam trap
The trap here is that candidates often confuse field attributes (searchable, retrievable) with the vector-specific configuration required for similarity scoring, leading them to pick options A or D instead of recognizing the missing vector configuration.
How to eliminate wrong answers
Option A is wrong because the 'searchable' attribute is not applicable to vector fields; vector fields are inherently searchable via vector queries, and setting them as searchable is not required for relevance sorting. Option B is wrong because mismatched vector dimensions would cause query failures or errors, not unsorted results; the question states the query runs but results are not sorted by relevance. Option D is wrong because the 'retrievable' attribute controls whether field values are returned in search results, not how results are sorted; setting it to false would omit the field from output but not affect relevance sorting.
Which THREE components are required to deploy a bot using Azure AI Language's conversational language understanding (CLU) and Azure Bot Service?
Select 3 answers
A.An Azure Bot Service bot with the CLU model as a skill.
B.A CLU project and a trained model.
C.A QnA Maker resource for FAQs.
D.A Translator resource for multilingual support.
E.A Microsoft Entra ID app registration for the bot.
AnswersA, B, E
The bot interacts with users and uses the CLU model.
Why this answer
Azure Bot Service can integrate a CLU model as a skill, enabling the bot to route user intents to the CLU service for natural language understanding. This is done by configuring the bot's skill endpoint to point to the CLU project's prediction URL, allowing the bot to leverage the trained model for intent classification and entity extraction.
Exam trap
The trap here is that candidates often assume QnA Maker or Translator are mandatory for a CLU-based bot, but CLU itself provides intent recognition and can be extended with other services only as needed, not as prerequisites.
A developer wants to integrate a pre-built AI model that can extract key information from invoices, such as vendor name, invoice date, and total amount. Which Azure AI service should they use?
A.Azure AI Language
B.Azure OpenAI Service
C.Azure AI Document Intelligence
D.Azure Cognitive Search
AnswerC
Document Intelligence provides pre-built invoice models for extraction.
Why this answer
Azure AI Document Intelligence (formerly Form Recognizer) is the correct service because it is specifically designed to extract structured data from documents like invoices, including fields such as vendor name, invoice date, and total amount. It uses pre-built models trained on common document types, making it ideal for this use case without requiring custom training.
Exam trap
The trap here is that candidates may confuse Azure AI Language's entity extraction capabilities with document-specific extraction, not realizing that Azure AI Document Intelligence is purpose-built for semi-structured documents like invoices and forms.
How to eliminate wrong answers
Option A is wrong because Azure AI Language focuses on text analytics, sentiment analysis, and entity recognition from unstructured text, not on extracting structured fields from semi-structured documents like invoices. Option B is wrong because Azure OpenAI Service provides generative AI models (e.g., GPT-4) for text generation and conversation, not a pre-built model optimized for invoice data extraction. Option D is wrong because Azure Cognitive Search is a search and indexing service for building search experiences over data, not a document processing service for extracting key information from invoices.
Refer to the exhibit. You are deploying a generative AI model as an online endpoint in Azure Machine Learning. You receive complaints that the endpoint returns 503 errors during peak hours. What is the most likely cause?
A.The manual scale setting with 2 instances may be insufficient for peak traffic.
B.The request timeout of 30 seconds is too short.
C.The environment variable MODEL_CACHE_SIZE is set too low.
D.The model version is not specified correctly.
AnswerA
Manual scaling does not automatically adjust; if traffic exceeds capacity, requests are rejected with 503.
Why this answer
503 errors during peak hours indicate that the endpoint is overwhelmed by the request volume. With manual scaling set to only 2 instances, the compute capacity is insufficient to handle the increased traffic, causing the service to reject requests. Azure Machine Learning online endpoints require sufficient instance count or autoscaling to absorb traffic spikes.
Exam trap
Microsoft often tests the distinction between HTTP status codes (503 vs. 408/504) to mislead candidates into confusing timeout-related errors with capacity-related errors.
How to eliminate wrong answers
Option B is wrong because a 30-second request timeout would cause timeout errors (e.g., 408 or 504), not 503 Service Unavailable errors; 503 specifically signals resource exhaustion, not slow processing. Option C is wrong because the MODEL_CACHE_SIZE environment variable affects model loading performance and caching, not the endpoint's ability to handle concurrent requests; a low cache size might cause slower inference but not 503 errors. Option D is wrong because an incorrect model version would result in deployment failures (e.g., 404 or 500 errors) during endpoint creation or update, not intermittent 503 errors during peak traffic.
You are deploying a custom question-answering solution using Azure AI Language. The solution must support multiple languages and provide answers from a set of FAQ documents. You need to ensure that the service can handle up to 1000 requests per minute with low latency. What should you configure?
A.Use Azure AI Search with a basic tier and replicas.
B.Use the Standard tier and allocate 10 capacity units (CUs).
C.Use the Free tier and enable autoscaling.
D.Deploy a dedicated endpoint with provisioned throughput of 1000 TPS.
AnswerB
Standard tier with CUs provides the required throughput.
Why this answer
Azure AI Language's custom question-answering feature uses a Standard tier with capacity units (CUs) to handle throughput and latency. Each CU provides a baseline of 100 requests per minute (RPM), so 10 CUs deliver exactly 1000 RPM, meeting the requirement. The Standard tier also supports multiple languages and low-latency responses from FAQ documents.
Exam trap
The trap here is confusing throughput units (requests per minute vs. per second) and assuming that Azure AI Search replicas or provisioned TPS from other services apply to Azure AI Language custom question-answering, when the correct scaling mechanism is capacity units (CUs) on the Standard tier.
How to eliminate wrong answers
Option A is wrong because Azure AI Search with a basic tier and replicas is designed for indexing and search, not for the managed question-answering runtime of Azure AI Language; replicas improve search query throughput but do not directly control the custom QA endpoint's request rate or latency. Option C is wrong because the Free tier is limited to 3 transactions per minute and cannot scale to 1000 requests per minute, and autoscaling is not available on the Free tier. Option D is wrong because provisioned throughput of 1000 TPS (transactions per second) is a concept for Azure OpenAI or dedicated endpoints in other services, not for Azure AI Language custom question-answering, which uses capacity units (CUs) measured in requests per minute, not per second.
A company uses Azure Custom Vision to build a classifier for defect detection on a manufacturing line. They have labeled images of products with and without defects. Which TWO actions should they take to improve model performance?
Select 2 answers
A.Train for more iterations without validation.
B.Use images with balanced numbers of defect and non-defect samples.
C.Set the learning rate manually using the Custom Vision API.
D.Increase the number of images per tag, including variations in lighting and angle.
E.Reduce the number of images per tag to avoid overfitting.
Balanced datasets prevent the model from becoming biased toward the majority class (e.g., non-defect images), which is critical for defect detection where defects are rare. Azure Custom Vision uses a weighted loss function during training, and class imbalance can cause the model to predict the majority class for most inputs, reducing recall for defects. Balanced samples ensure the model learns discriminative features for both classes equally.
Exam trap
The trap here is that candidates may think reducing images prevents overfitting (Option E) or that manual learning rate tuning (Option C) is possible in Custom Vision, but the service abstracts hyperparameter tuning and requires sufficient, varied data for robust defect detection.
Your knowledge mining solution ingests documents from multiple tenants. Each tenant's data must be isolated and searchable only by that tenant. You have a single Azure AI Search service. How should you implement multi-tenancy?
A.Use separate skillsets for each tenant
B.Create a separate search service for each tenant
C.Use a single index with a tenant ID field and filter queries by that field
D.Use separate data sources within the same index
AnswerC
Index-level security with filters is the recommended approach.
Why this answer
Azure AI Search supports multi-tenancy within a single service by using a shared index with a tenant ID field. Each document is tagged with a tenant identifier, and queries are scoped using OData `$filter` expressions (e.g., `$filter=tenantId eq 'tenant123'`). This ensures data isolation while keeping costs low and management simple, as only one search service and one index are needed.
Exam trap
The trap here is that candidates confuse data sources (which are just ingestion pipelines) with data partitioning, leading them to think separate data sources or skillsets provide isolation, when in fact only query-time filtering or separate indexes enforce tenant boundaries.
How to eliminate wrong answers
Option A is wrong because skillsets define enrichment pipelines (e.g., OCR, entity extraction) and are not used for data isolation; they apply to all documents in an index regardless of tenant. Option B is wrong because creating a separate search service for each tenant is unnecessarily expensive and complex, violating the requirement to use a single Azure AI Search service. Option D is wrong because separate data sources within the same index still store all documents together; data sources only define where data is pulled from, not how it is partitioned or secured at query time.
You have defined the custom WebApiSkill shown in the exhibit. The skill calls an Azure Function that can process up to 10 documents per second. However, you notice that the skill is failing with 429 errors. What is the most likely cause?
A.The timeout of 30 seconds is too short for the function to respond
B.The batch size of 5 is too large, causing the function to receive too many documents at once
C.The context '/document' is incorrect, causing all documents to be processed as one
D.The degreeOfParallelism of 3 causes too many concurrent requests, exceeding the function's capacity
AnswerD
With batchSize 5 and degreeOfParallelism 3, up to 15 documents are sent concurrently.
Why this answer
The `degreeOfParallelism` of 3 causes the AI Search enrichment pipeline to invoke the Azure Function with up to 3 concurrent batches, each of size 5, resulting in up to 15 documents per second. Since the function can only handle 10 documents per second, this exceeds its capacity and triggers HTTP 429 (Too Many Requests) errors.
Exam trap
The trap here is that candidates often focus on the batch size as the sole cause of rate limiting, overlooking that `degreeOfParallelism` multiplies the effective request rate, which is the actual trigger for 429 errors.
How to eliminate wrong answers
Option A is wrong because a 30-second timeout is typically sufficient for an Azure Function processing documents; 429 errors indicate rate limiting, not timeout. Option B is wrong because a batch size of 5 means the function receives 5 documents per invocation, which is within the 10-document-per-second capacity if only one batch is processed at a time. Option C is wrong because the context '/document' is the standard path for per-document processing in AI Search skills; using it does not cause all documents to be processed as one, but rather each document is processed individually.
Which TWO Azure AI services can be used together to build a solution that transcribes customer service calls and detects sentiment?
Select 2 answers
A.Azure AI Language (sentiment analysis)
B.Azure AI Speech (speech-to-text)
C.Azure AI Translator
D.Azure AI Speech (text-to-speech)
E.Azure AI Language (conversational language understanding)
AnswersA, B
Analyzes sentiment from text.
Why this answer
Azure AI Speech's speech-to-text capability transcribes the audio of customer service calls into text, while Azure AI Language's sentiment analysis evaluates that text to determine the overall sentiment (e.g., positive, negative, or neutral). Together, they form a complete pipeline for analyzing call recordings. Option A is correct because sentiment analysis is the specific feature within Azure AI Language that detects sentiment from text.
Option B is correct because speech-to-text is the specific feature within Azure AI Speech that converts audio to text.
Exam trap
The trap here is that candidates may confuse Azure AI Speech's text-to-speech with speech-to-text, or mistakenly think Azure AI Translator or conversational language understanding can perform sentiment analysis, when in fact only the specific sentiment analysis feature of Azure AI Language is designed for that task.
You deployed a custom neural voice in Azure AI Speech. The model generates poor prosody for long sentences. What should you do?
A.Re-record the training audio at 24 kHz sample rate.
B.Switch to a prebuilt neural voice.
C.Increase the number of training scripts.
D.Use SSML tags to control prosody at runtime.
AnswerA
High-quality audio is essential for learning natural prosody.
Why this answer
A custom neural voice model requires high-quality training data. A 24 kHz sample rate is the standard for Azure AI Speech's neural voices, as it captures the full frequency range needed for natural prosody. Re-recording at this rate ensures the model learns proper intonation and rhythm for long sentences, directly addressing the poor prosody issue.
Exam trap
The trap here is that candidates may think runtime SSML tags (Option D) can fix poor prosody, but the question specifies the model generates poor prosody, indicating a training data quality issue that SSML cannot correct.
How to eliminate wrong answers
Option B is wrong because switching to a prebuilt neural voice does not fix the custom model's training data quality issue; it only replaces the model entirely, which may not meet custom voice requirements. Option C is wrong because increasing the number of training scripts does not improve prosody if the existing audio quality (e.g., sample rate) is inadequate; more low-quality data does not help. Option D is wrong because SSML tags control prosody at runtime but do not address the root cause of poor prosody in the trained model; they are a workaround, not a fix for training data deficiencies.
You are designing a solution that uses Azure AI Translator to translate documents in real-time. The solution must handle up to 10,000 characters per request and must be cost-effective. Which pricing tier should you select?
A.F0 (Free)
B.S1 (Standard)
C.D3 (Volume)
D.S2 (Standard)
AnswerB
S1 supports up to 10,000 characters per request at standard cost.
Why this answer
The S1 (Standard) tier is correct because it supports up to 10,000 characters per request for real-time document translation, which meets the requirement, and it is the most cost-effective paid tier for this workload. The Free tier (F0) has a limit of 5,000 characters per request and is not suitable for production use, while higher tiers like S2 and D3 are unnecessary and more expensive for this character limit.
Exam trap
The trap here is that candidates often confuse the character-per-request limit with the total monthly character allowance, leading them to incorrectly select the Free tier (F0) for small workloads without realizing its 5,000-character per request cap is insufficient for the stated requirement.
How to eliminate wrong answers
Option A is wrong because the F0 (Free) tier has a maximum of 5,000 characters per request, which is insufficient for handling 10,000 characters per request, and it is intended only for evaluation and non-production use. Option C is wrong because the D3 (Volume) tier is designed for high-volume translation with discounted per-character pricing but is not cost-effective for the specified 10,000-character limit, as it requires a commitment to large monthly volumes and has a higher base cost. Option D is wrong because the S2 (Standard) tier supports the same 10,000-character limit as S1 but at a higher per-character cost, making it less cost-effective for this requirement.
A manufacturing company uses Azure AI Custom Vision to classify defects on assembly line parts. The model is currently trained with 500 images per class and achieves 85% accuracy. The company needs to improve accuracy to 95% without collecting new images. Which THREE actions should they take?
Select 3 answers
A.Change the domain to 'General'
B.Use data augmentation (e.g., rotation, scaling)
C.Increase the number of training iterations
D.Adjust the probability threshold
E.Add negative images (images without defects)
AnswersB, D, E
Creates variations from existing images.
Why this answer
Custom Vision's built-in data augmentation (rotation, scaling, flipping, etc.) artificially increases the diversity of the training set without requiring new images. This helps the model generalize better and reduces overfitting, directly improving accuracy toward the 95% target.
Exam trap
The trap here is that candidates often confuse adjusting the probability threshold (a post-processing step) with improving model accuracy, or mistakenly think increasing iterations always helps, when in fact Custom Vision's training is already optimized to converge automatically.
You are a data scientist at a healthcare research organization. You have been tasked with building a knowledge mining solution to extract key information from thousands of medical journal articles stored as PDFs in an Azure Blob Storage container. The articles are in English and contain tables, figures, and structured text. Your organization uses Microsoft Purview for data governance. You need to design a solution that uses Azure AI Search and Azure AI Services to extract and index the following: article title, authors, publication date, abstract, and key findings (as key phrases). The solution must also detect any mentions of drugs and dosages. The extracted information must be indexed and searchable via a custom web application. Which approach should you take?
A.Use Azure AI Search with a skillset that includes OCR skill, Text Translation skill to translate, Entity Recognition skill for drugs, and Key Phrase Extraction. Index the results.
B.Use Azure AI Search with a blob indexer that includes a skillset with Document Layout skill to extract text, Key Phrase Extraction skill to extract key findings, and map built-in metadata for title, authors, date. Use a custom index to store the extracted fields.
C.Use Azure AI Search with a blob indexer and a skillset that includes Document Layout skill, Entity Recognition skill to extract drug names, Key Phrase Extraction skill, and custom skill to extract title/authors/date from the first page. Create an index with fields for each required element.
D.Use Azure AI Search with a skillset that includes OCR skill, Entity Recognition skill, Sentiment skill, and Key Phrase Extraction. Use a knowledge store to project the enriched data.
AnswerC
Covers all requirements with appropriate skills.
Why this answer
It combines the Document Layout skill (to extract text from PDFs including tables and figures), Entity Recognition skill (to detect drug names), Key Phrase Extraction skill (to identify key findings), and a custom skill (to parse the first page for title, authors, and publication date). This approach handles the unstructured nature of medical journal articles while meeting all extraction requirements and indexing them into a custom searchable index.
Exam trap
The trap here is that candidates often assume built-in blob metadata (like 'metadata_storage_name') can extract article-specific fields like authors or publication date, but those are only file-level properties and require a custom skill to parse from the document content.
How to eliminate wrong answers
Option A is wrong because it uses OCR skill (unnecessary for digital PDFs with selectable text) and Text Translation skill (not needed since articles are in English), and it lacks a custom skill to extract title, authors, and date from the first page. Option B is wrong because it relies on built-in metadata for title, authors, and date, but blob indexer metadata only captures file-level properties (e.g., filename, size) and cannot extract article-specific fields like authors or publication date from within the document. Option D is wrong because it includes Sentiment skill (irrelevant for extracting structured information) and uses a knowledge store instead of directly indexing the enriched data for search, and it omits the Document Layout skill and custom skill needed for title/authors/date extraction.
Refer to the exhibit. You are using Azure OpenAI to build a trivia bot. The bot responds correctly to simple questions but sometimes gives overly verbose or off-topic answers. Which parameter should you adjust to make responses more focused?
A.Increase temperature
B.Decrease top_p
C.Increase max_tokens
D.Decrease max_tokens
E.Decrease temperature
AnswerE
Lower temperature makes the model more deterministic and focused.
Why this answer
Decreasing the temperature parameter reduces the randomness of the model's output, making it more deterministic and focused. For a trivia bot that gives overly verbose or off-topic answers, lowering temperature (e.g., from 0.7 to 0.3) forces the model to choose higher-probability tokens, which typically correspond to more concise and relevant responses.
Exam trap
Microsoft often tests the misconception that max_tokens controls response focus, but candidates must understand that max_tokens only limits length, not content relevance, while temperature directly influences the model's creativity and adherence to the prompt.
How to eliminate wrong answers
Option A is wrong because increasing temperature would increase randomness, making responses even more verbose and off-topic. Option B is wrong because decreasing top_p (nucleus sampling) can also reduce diversity but is not the primary parameter for controlling focus; it cuts off low-probability tokens, which may help but is less direct than temperature for this symptom. Option C is wrong because increasing max_tokens would allow longer responses, exacerbating verbosity.
Option D is wrong because decreasing max_tokens truncates the output length but does not address the core issue of off-topic content; the model may still produce irrelevant text within the shorter limit.
You are developing a knowledge mining solution that extracts insights from customer feedback. Which TWO Azure AI services can be used to analyze the sentiment of the feedback and categorize it into topics?
Select 2 answers
A.Azure AI Personalizer
B.Azure AI Translator
C.Azure AI Custom Vision
D.Azure AI Language
E.Azure AI Search with cognitive skills
AnswersD, E
Azure AI Language includes sentiment analysis and key phrase extraction for topic identification.
Why this answer
Azure AI Language provides pre-built sentiment analysis and key phrase extraction capabilities that can determine the sentiment (positive, negative, neutral) of customer feedback. Additionally, its named entity recognition and topic modeling features allow you to categorize feedback into topics, making it a direct fit for this requirement.
Exam trap
The trap here is that candidates often confuse Azure AI Language with Azure AI Translator or Azure AI Personalizer, mistakenly thinking translation or personalization services can perform sentiment analysis and topic categorization, when in fact only Azure AI Language provides those text analytics capabilities.
You are designing an agentic solution using Azure AI Agent Service with a custom skill that calls an external REST API. The API has rate limits: 100 requests per minute per client. You need to ensure the agent respects this limit without degrading user experience. Which approach should you take?
A.Use a token bucket rate limiter with a shared counter stored in Azure Cache for Redis
B.Set a fixed delay of 600ms between each API call
C.Configure the skill to handle HTTP 429 responses with retry-after logic
D.Implement exponential backoff in the skill code
AnswerA
Token bucket allows burst and respects limits globally.
Why this answer
A token bucket rate limiter with a shared counter stored in Azure Cache for Redis provides a distributed, atomic mechanism to enforce a strict 100 requests per minute per client limit across multiple instances of the agent. This approach ensures that the rate limit is respected globally without introducing unnecessary delays when the limit is not reached, thus maintaining a responsive user experience.
Exam trap
The trap here is that candidates often confuse reactive error-handling strategies (like handling 429 responses or exponential backoff) with proactive rate limiting, failing to recognize that only a distributed, preemptive mechanism like a token bucket with a shared counter can prevent rate limit violations without degrading user experience.
How to eliminate wrong answers
Option B is wrong because a fixed delay of 600ms between each API call (which would allow ~100 requests per minute) does not account for network latency, processing time, or concurrent calls from multiple agent instances, and it would artificially slow down all requests even when the rate limit is not approached, degrading user experience. Option C is wrong because handling HTTP 429 responses with retry-after logic is a reactive approach that only addresses rate limit violations after they occur, leading to failed requests and retries that degrade user experience, rather than proactively preventing the limit from being exceeded. Option D is wrong because implementing exponential backoff in the skill code is also a reactive strategy that waits for a 429 response before backing off, which still results in failed requests and increased latency, and it does not provide a shared counter to coordinate across distributed agent instances.
Your company wants to moderate user-uploaded images for adult content. Which Azure AI service should you use?
A.Azure AI Content Safety
B.Azure AI Face
C.Azure AI Document Intelligence
D.Azure AI Vision Image Analysis
AnswerA
Content Safety provides content moderation for images and text.
Why this answer
Azure AI Content Safety is the correct service because it is specifically designed to detect and moderate inappropriate content, including adult content, in images and text. It provides severity-based classifications (safe, low, medium, high) for categories such as hate, self-harm, sexual, and violence, making it ideal for user-uploaded image moderation.
Exam trap
The trap here is that candidates often confuse Azure AI Vision Image Analysis (which can detect adult content via the 'adult' flag in its Analyze Image API) with the dedicated Azure AI Content Safety service, but the exam expects the service purpose-built for content moderation with granular severity levels and broader category support.
How to eliminate wrong answers
Option B is wrong because Azure AI Face is focused on detecting, analyzing, and recognizing human faces, not on moderating adult content. Option C is wrong because Azure AI Document Intelligence is designed to extract text, key-value pairs, and tables from documents, not to analyze images for adult content. Option D is wrong because Azure AI Vision Image Analysis provides general image descriptions, object detection, and optical character recognition, but lacks the specific content moderation categories and severity scoring needed for adult content detection.
You deploy the ARM template shown in the exhibit to create an Azure AI Search indexer. The indexer fails to run, and you see an error that the skillset 'demo-skillset' does not exist. What is the most likely cause?
A.The field mapping source field 'metadata_storage_path' is incorrect
B.The schedule start time is in the past, causing the indexer to be disabled
C.The data source 'demo-datasource' does not exist
D.The skillset resource was not deployed before the indexer
AnswerD
The indexer depends on the skillset, which must exist. The template does not include the skillset resource.
Why this answer
The ARM template deploys resources in parallel by default, so the indexer can be created before the skillset. Since the indexer references 'demo-skillset' in its configuration, the deployment fails because the skillset resource does not exist at the time the indexer is created. To fix this, you must add a 'dependsOn' property in the indexer resource definition to ensure the skillset is deployed first.
Exam trap
The trap here is that candidates assume ARM templates deploy resources in the order they appear in the template, but Azure deploys them in parallel by default, so missing 'dependsOn' causes dependency errors like the one shown.
How to eliminate wrong answers
Option A is wrong because 'metadata_storage_path' is a valid system-generated field in Azure Blob Storage data sources, and an incorrect field mapping would cause a field mapping error, not a 'skillset does not exist' error. Option B is wrong because a schedule start time in the past does not disable the indexer; it simply means the first run is scheduled immediately, and the indexer remains active. Option C is wrong because the error message explicitly states the skillset does not exist, not the data source; if the data source were missing, the error would reference 'demo-datasource' instead.
Your company has a large set of PDF documents stored in Azure Blob Storage. You need to index these documents in Azure Cognitive Search so that users can search the text content. What is the first step you should take?
A.Create an index with a field for each metadata property.
B.Create a skillset to extract text from PDFs.
C.Create a data source that connects to Azure Blob Storage.
D.Create an indexer that runs daily.
AnswerC
A data source is required to specify where the data is located.
Why this answer
The correct first step is to create a data source that connects to Azure Blob Storage (option C). In Azure Cognitive Search, the indexing pipeline requires a data source definition to specify the location and type of data. After creating the data source, you can then create an index, and finally an indexer to automate the process.
Creating a skillset (option B) or an indexer running daily (option D) would come later. Option A is incorrect because the index fields are defined when you create the index, which is not the first step.
You are developing a chat application that uses Azure OpenAI Service to answer customer queries. The solution must ensure that the model does not generate harmful or offensive content. Which Azure AI service should you configure?
A.Azure AI Bot Service
B.Azure AI Search
C.Azure AI Content Safety
D.Azure AI Language
AnswerC
Azure AI Content Safety provides content moderation to filter harmful content.
Why this answer
Azure AI Content Safety is specifically designed to detect and filter harmful or offensive content in text and images, making it the appropriate service to integrate with an Azure OpenAI chat application to enforce content safety policies. It provides APIs for content moderation, including severity-based filtering for hate, self-harm, sexual, and violence categories, which directly addresses the requirement to prevent the model from generating harmful output.
Exam trap
The trap here is that candidates often confuse Azure AI Language's text analytics capabilities (like sentiment analysis) with content safety, assuming that language understanding inherently includes harm detection, but Azure AI Content Safety is a separate, specialized service for content moderation.
How to eliminate wrong answers
Option A is wrong because Azure AI Bot Service is a platform for building, testing, and deploying conversational agents (bots), not a content moderation or safety service; it does not natively filter harmful content from model outputs. Option B is wrong because Azure AI Search is a search-as-a-service solution for indexing and querying data, with no built-in content safety or moderation capabilities for generated text. Option D is wrong because Azure AI Language provides natural language processing features like sentiment analysis, key phrase extraction, and question answering, but it does not include content safety filters for detecting harmful or offensive content.
You run the above PowerShell command to check the responsible AI policy applied to an Azure OpenAI Service deployment. The output shows 'MyPolicy'. You need to verify that the policy blocks hate speech. What should you do?
A.Use the Azure Portal to review the deployment's properties.
B.Navigate to Azure AI Content Safety and view the policy 'MyPolicy'.
C.Run Get-AzCognitiveServicesAccountDeployment with -ExpandProperties.
D.Use Azure Monitor to check the deployment's logs.
AnswerB
Content Safety is where policies are defined and can be viewed.
Why this answer
Azure AI Content Safety is the dedicated service for managing and viewing the details of content filtering policies, including custom policies like 'MyPolicy'. The PowerShell command only confirms the policy name is applied to the deployment; to verify that the policy blocks hate speech, you must inspect the policy's configuration (e.g., severity thresholds for hate categories) within the Azure AI Content Safety portal. The Azure Portal deployment properties only show the policy name, not its rules.
Exam trap
The trap here is that candidates assume the deployment properties or PowerShell output contain the policy's rule details, when in fact they only show the policy name, leading them to choose Option A or C instead of navigating to the dedicated Content Safety service where the actual filter rules are configured.
How to eliminate wrong answers
Option A is wrong because the Azure Portal's deployment properties only display the policy name (e.g., 'MyPolicy') and do not expose the specific content filter rules or severity thresholds that define what is blocked. Option C is wrong because Get-AzCognitiveServicesAccountDeployment with -ExpandProperties returns deployment-level metadata (e.g., model, scale settings) but not the content filter policy's rule details; it cannot show whether hate speech is blocked. Option D is wrong because Azure Monitor logs capture operational events and usage metrics, not the configuration of content filtering policies; logs cannot reveal the policy's rule definitions.
Refer to the exhibit. You run the Azure CLI command shown. What is the output?
A.A single key string
B.Both key1 and key2
C.The endpoint URL
D.The region of the account
AnswerA
The query extracts key1 as a string.
Why this answer
The `az cognitiveservices account keys list` command retrieves the access keys for an Azure Cognitive Services account. By default, it returns both key1 and key2, but the `--query` parameter with `keys[0]` filters the output to return only the first key (key1) as a single string. This is a standard Azure CLI query syntax using JMESPath to extract a specific element from the JSON response.
Exam trap
Microsoft often tests the distinction between retrieving keys versus other account properties (like endpoint or region), and the trap here is that candidates assume `keys list` returns all keys by default without recognizing that the `--query` parameter can filter to a single key.
How to eliminate wrong answers
Option B is wrong because the `--query keys[0]` filter explicitly selects only the first key (key1), not both key1 and key2. Option C is wrong because the command retrieves keys, not the endpoint URL; the endpoint URL is obtained via `az cognitiveservices account show` with a different query. Option D is wrong because the region is a property of the account resource, not the keys list; the region is returned by `az cognitiveservices account show` or `list`, not by the keys list command.
Which TWO features are available in Azure Computer Vision but not in Azure Custom Vision?
Select 2 answers
A.Object detection
B.Image classification
C.Face detection
D.Optical character recognition (OCR)
E.Image captioning
AnswersD, E
Pre-built OCR only in Computer Vision.
Why this answer
Optical Character Recognition (OCR) and image captioning are features available in Azure Computer Vision but not in Azure Custom Vision. Azure Computer Vision provides pre-built OCR APIs (e.g., Read API) for extracting text from images and captioning models for generating human-readable descriptions. Azure Custom Vision is focused on training custom models for image classification and object detection, and does not include built-in OCR or captioning capabilities.
Exam trap
The trap here is that candidates often assume all computer vision features are available in both services, but Azure Custom Vision is specifically designed for custom model training and does not include pre-built OCR or image captioning APIs, which are exclusive to Azure Computer Vision.
You are a data scientist for Contoso Pharmaceuticals. The company has thousands of research documents in PDF format stored in Azure Blob Storage. You need to build an Azure Cognitive Search solution that enables researchers to search for documents based on chemical compound names, disease mentions, and experimental results. The solution must extract these entities using a custom AI model built in Azure AI Language. Additionally, the solution must support semantic search for natural language queries. The search index must be updated daily with new documents. You have an existing Azure AI Language custom entity extraction model that recognizes chemical compounds and diseases. The model is deployed as an endpoint. You need to configure the enrichment pipeline. What should you do?
A.Create a custom skill in the skillset that calls the custom entity extraction endpoint via HTTP.
B.Deploy the custom model to Azure AI Document Intelligence and use a Document Intelligence skill.
C.Add the custom entity extraction as a field mapping in the indexer.
D.Use the built-in Entity Recognition skill and configure it to use your custom model endpoint.
AnswerA
Custom skills can call external APIs, including custom model endpoints.
Why this answer
To integrate a custom AI model from Azure AI Language into an Azure Cognitive Search enrichment pipeline, you need to create a custom skill in the skillset that calls the custom entity extraction endpoint via HTTP. The built-in Entity Recognition skill only supports prebuilt models and cannot be configured to use a custom model endpoint. Deploying the model to Azure AI Document Intelligence and using a Document Intelligence skill is not appropriate because the model is already deployed in Azure AI Language as a custom entity extraction model.
Field mappings in the indexer are for direct field-to-field mappings from the data source to the index, not for calling external AI services. Therefore, Option A is the correct approach.
You need to extract personally identifiable information (PII) from a set of text documents before indexing them in Azure AI Search. The PII must be redacted. Which Azure AI service and configuration should you use?
A.Use the entity recognition skill in Azure AI Search and map to a target field
B.Use Azure AI Document Intelligence with a custom model to identify PII fields
C.Use the built-in PII detection skill in Azure AI Search with redaction mode enabled
D.Use Azure AI Language's key phrase extraction to find PII
AnswerC
The PII detection skill can redact detected entities.
Why this answer
Azure AI Search provides a built-in PII detection skill that can automatically identify and redact PII entities (such as names, phone numbers, and email addresses) from text documents during the indexing pipeline. The skill supports a 'redactionMode' configuration that replaces detected PII with a placeholder (e.g., '***'), meeting the requirement to redact PII before indexing.
Exam trap
The trap here is that candidates confuse the general entity recognition skill (which identifies entities but does not redact) with the PII-specific skill (which is designed for redaction), leading them to choose Option A instead of C.
How to eliminate wrong answers
Option A is wrong because the entity recognition skill in Azure AI Search identifies entities (e.g., people, organizations) but does not natively support PII-specific redaction; it outputs entity categories and values, not redacted text. Option B is wrong because Azure AI Document Intelligence with a custom model is designed for extracting structured data from forms and documents, not for general-purpose PII detection and redaction from arbitrary text. Option D is wrong because Azure AI Language's key phrase extraction identifies key topics and phrases, not PII entities, and lacks any redaction capability.
You are using Azure AI Document Intelligence to extract data from scanned contracts. The contracts contain tables and handwritten signatures. Which TWO features should you enable?
Select 2 answers
A.Train a custom neural model to recognize handwritten signatures.
B.Enable table extraction in the custom model.
C.Enable OCR to read scanned text.
D.Use form recognition to capture key-value pairs.
E.Use the prebuilt-layout model for all extraction.
AnswersA, B
Neural models can learn to extract signatures.
Why this answer
Options A and B are correct because a custom neural model can be trained to recognize handwritten signatures (A), and enabling table extraction in the custom model extracts tables from contracts (B). Option C is incorrect because OCR is already enabled by default in Document Intelligence and does not need to be separately enabled. Option D is wrong because form recognition for key-value pairs is not required for this scenario, as the focus is on tables and signatures.
Option E is wrong because the prebuilt-layout model does not support custom training for signature recognition and may not handle handwritten elements as accurately as a custom model.
You are building a generative AI application that uses Azure OpenAI Service. The application must access data from an Azure SQL database to answer user questions. You need to implement a solution that retrieves the most relevant data without exposing the database schema to the model. Which approach should you use?
A.Configure the model to generate SQL queries and execute them against the database
B.Use a middle-tier service to query the database, then pass results to the model via Azure OpenAI On Your Data
C.Embed the entire database schema in the system prompt
D.Fine-tune the model with historical query results from the database
AnswerB
This isolates the database and provides relevant context to the model.
Why this answer
It uses a middle-tier service to query the Azure SQL database and then passes only the relevant results to the Azure OpenAI model via the 'On Your Data' feature. This approach ensures the model never sees the database schema, protecting sensitive structural details while still providing the necessary context for answering user questions. It also allows the middle-tier to handle authentication, query optimization, and data filtering securely.
Exam trap
The trap here is that candidates often assume the model can safely generate SQL queries (Option A) because it seems efficient, but they overlook the security and schema exposure risks, as well as the fact that Azure OpenAI On Your Data is the designed pattern for this exact use case.
How to eliminate wrong answers
Option A is wrong because having the model generate SQL queries directly exposes the database schema to the model (either explicitly or implicitly through query patterns) and introduces significant security risks, such as SQL injection or unintended data access, which violates the requirement to not expose the schema. Option C is wrong because embedding the entire database schema in the system prompt would directly expose the schema to the model, contradicting the requirement, and would also consume excessive token limits, potentially degrading performance. Option D is wrong because fine-tuning the model with historical query results does not address the need to retrieve the most relevant data dynamically; it only teaches the model to mimic past responses without providing a mechanism to query the live database, and it still risks exposing schema information through the training data.
You are building an agent in Microsoft Copilot Studio that needs to send a confirmation email after a user completes a survey. The email should be sent using the user's email address collected during the conversation. Which feature should you use to send the email?
A.Create a Power Automate flow that sends an email and call it from the topic.
B.Use the Send an email action directly in Copilot Studio.
C.Use an adaptive card with an email button.
D.Use the email channel to send a response.
AnswerA
Power Automate flows can send emails via connectors like Outlook.
Why this answer
Microsoft Copilot Studio does not have a native 'Send an email' action; it relies on Power Automate flows to perform external actions like sending emails. By creating a flow that uses the user's email address collected during the conversation and calling it from the topic, you can send a confirmation email after the survey is completed.
Exam trap
The trap here is that candidates assume Copilot Studio has built-in email actions similar to Power Automate, but Microsoft deliberately separates conversation logic from external integrations to enforce a modular architecture.
How to eliminate wrong answers
Option B is wrong because Copilot Studio does not include a built-in 'Send an email' action; it lacks native email capabilities and must delegate such tasks to Power Automate. Option C is wrong because an adaptive card with an email button only provides a clickable interface to open the user's default mail client; it does not programmatically send an email from the bot. Option D is wrong because the email channel in Copilot Studio is used to receive and respond to messages via email, not to send outbound emails like a confirmation.
Your knowledge mining solution uses Azure AI Search. Users complain that search results are not relevant. You have enabled semantic search but results still lack context. What should you do to improve relevance?
A.Ensure the index includes a semantic configuration with title and content fields
B.Increase the number of partitions to handle more data
C.Configure a scoring profile with boosting based on metadata
D.Increase the number of replicas to improve query performance
AnswerA
Semantic configuration is required for semantic ranking to work.
Why this answer
Semantic search in Azure AI Search requires a semantic configuration that explicitly maps the title and content fields to be used for semantic ranking. Without this configuration, the search engine cannot apply the deep neural network models that understand context and intent, so results remain based on keyword matching even when the semantic search feature is enabled.
Exam trap
The trap here is that candidates assume enabling the semantic search feature alone is sufficient, but Azure AI Search requires an explicit semantic configuration to map the fields that the semantic model will use for reranking.
How to eliminate wrong answers
Option B is wrong because increasing the number of partitions scales the index for larger data volumes but does not improve relevance or semantic understanding of search results. Option C is wrong because scoring profiles with boosting based on metadata can adjust ranking weights but do not provide the contextual, language-understanding capabilities that semantic search offers. Option D is wrong because increasing replicas improves query throughput and availability, not the relevance or contextual quality of search results.
You have the above skillset in Azure AI Search. The indexer processes a document with 12,000 characters of content. How many entity recognition skill executions occur?
A.4
B.2
C.3
D.1
AnswerC
Three pages result from the split, each triggering an entity recognition execution.
Why this answer
The Azure AI Search entity recognition skill has a maximum text length per execution of 5,000 characters. A document with 12,000 characters is split into chunks of up to 5,000 characters, resulting in three chunks (5,000 + 5,000 + 2,000). Each chunk triggers one skill execution, so three executions occur.
Exam trap
The Azure AI Search entity recognition skill has a maximum text length per execution of 5,000 characters. Candidates might incorrectly divide 12,000 by 5,000 and round down to 2, or assume a single execution can handle the entire document, ignoring the chunking behavior.
How to eliminate wrong answers
Option A is wrong because 4 executions would require more than 15,000 characters (4 × 5,000), but the document has only 12,000 characters. Option B is wrong because 2 executions would cover only 10,000 characters (2 × 5,000), leaving 2,000 characters unprocessed. Option D is wrong because 1 execution can handle only up to 5,000 characters, but the document has 12,000 characters, so it must be split into multiple chunks.
A news organization uses Azure Video Indexer to generate transcripts of live broadcasts. They notice that the speaker names are not appearing in the transcript. What is the most likely cause?
A.The video resolution is too low for OCR.
B.The speaker identification model has not been trained with voice samples.
Speaker names are missing because Azure Video Indexer's speaker identification feature requires pre-trained voice samples to match speakers to their identities. Without a custom voice model trained on known speakers' audio, the service can only label speakers as 'Speaker #1', 'Speaker #2', etc., but cannot assign actual names. This is a supervised learning process where the model must be trained with labeled voice samples before it can recognize and name speakers.
Exam trap
The trap here is that candidates may confuse speaker identification with automatic diarization or assume that speaker names are automatically extracted from the video metadata, when in fact Azure Video Indexer requires explicit training of a custom Person Model with voice samples to assign names.
How to eliminate wrong answers
Option A is wrong because OCR (optical character recognition) is used for extracting text from video frames, not for identifying speakers or generating transcripts; low resolution would affect text extraction but not speaker name assignment. Option C is wrong because Azure Video Indexer supports a wide range of common video formats (e.g., MP4, MOV, AVI), and an unsupported format would cause a failure to index or generate any transcript, not just missing speaker names. Option D is wrong because setting the language incorrectly would affect the accuracy of the speech-to-text transcription (e.g., wrong words or gibberish), but it would not prevent speaker names from appearing; speaker identification is a separate model that requires training regardless of language.
You see the exhibit in an Azure AI Search skillset definition. The custom skill is not executing as expected. What is the most likely cause?
A.The skill is missing the 'uri' property to specify the endpoint
B.The input source path is incorrect
C.The JSON is malformed and will not parse
D.The type should be 'Microsoft.Skills.Custom.WebApiSkill' with dots
AnswerA
A custom web API skill must include 'uri' for the function URL.
Why this answer
The custom skill is not executing because the 'uri' property, which specifies the endpoint URL for the web API, is missing from the skill definition. Azure AI Search requires this property to route requests to the external service; without it, the skill cannot be invoked and will fail silently or return errors.
Exam trap
The trap here is that candidates often focus on input/output paths or syntax errors, overlooking the fact that the 'uri' property is a required field for custom WebApiSkills, and its absence is a common oversight in skillset definitions.
How to eliminate wrong answers
Option B is wrong because an incorrect input source path would cause a different symptom—such as missing or null data being passed to the skill—but the skill would still attempt execution; the core issue here is the missing endpoint. Option C is wrong because if the JSON were malformed, the skillset definition would fail to parse entirely during creation or update, not just cause the skill to not execute. Option D is wrong because the type 'Microsoft.Skills.Custom.WebApiSkill' is the correct and required format for a custom web API skill in Azure AI Search; using dots is standard and not a problem.
You have created a Copilot Studio agent that uses a Power Automate flow to send an email when a user requests a password reset. The flow works correctly in test mode, but when the agent is published, the flow does not run. What should you check first?
A.Check if the flow is shared with the Copilot Studio agent.
B.Check if the user has permission to run the flow.
C.Check if the trigger condition is correct.
D.Check if the flow is turned on.
AnswerA
Flows used in Copilot Studio must be shared with the agent's owner or the bot application.
Why this answer
When a Copilot Studio agent uses a Power Automate flow, the flow must be explicitly shared with the agent (the agent's application registration) so that the agent can trigger the flow on behalf of the user. In test mode, the flow runs under the creator's identity, but in published mode, the agent runs under its own identity, which requires the flow to be shared with the agent's service principal. Option A is correct because this is the most common cause of a flow working in test but failing after publication.
Exam trap
The trap here is that candidates assume the flow's trigger condition or 'turned on' status is the issue, overlooking the critical identity and permission handoff between the agent and the flow in published mode.
How to eliminate wrong answers
Option B is wrong because the flow runs under the agent's identity, not the end user's identity; checking user permissions is irrelevant unless the flow itself accesses a resource that requires user delegation. Option C is wrong because the trigger condition is already verified to work in test mode, so the issue is not with the trigger logic but with the runtime identity. Option D is wrong because the flow is confirmed to be turned on (it works in test mode), and the problem is specifically about the agent's ability to invoke the flow after publication.
Your company uses Microsoft Copilot for Microsoft 365. You need to ensure that Copilot only accesses data from approved SharePoint sites and does not use any other organizational data. What should you configure?
A.Configure Conditional Access policies in Microsoft Entra ID.
B.Use Microsoft Purview Data Map to catalog the approved sites.
C.Apply sensitivity labels to the approved SharePoint sites.
D.Set up data retention policies in Microsoft Purview.
AnswerC
Sensitivity labels can be used to restrict Copilot's data sources.
Why this answer
Sensitivity labels can be configured to restrict Microsoft Copilot for Microsoft 365 to only access content from approved SharePoint sites. By applying a sensitivity label that includes the 'mark content' or 'encrypt content' setting with a specific scope, you can use the 'Microsoft Copilot for Microsoft 365' condition under 'Access control' to block Copilot from processing data from unlabeled or non-approved sites. This ensures Copilot respects the label's policy and excludes all other organizational data.
Exam trap
The trap here is that candidates often confuse data access control with authentication or data lifecycle management, leading them to select Conditional Access or retention policies instead of recognizing that sensitivity labels provide the specific Copilot data source restriction capability.
How to eliminate wrong answers
Option A is wrong because Conditional Access policies in Microsoft Entra ID control user authentication and access to applications, not the data sources that Copilot can index or retrieve content from. Option B is wrong because Microsoft Purview Data Map is a metadata catalog for data governance and discovery, not a mechanism to enforce access restrictions on Copilot's data sources. Option D is wrong because data retention policies in Microsoft Purview manage how long data is kept and when it is deleted, not which data Copilot is allowed to access.
You have configured a system message for an Azure OpenAI chat completion deployment as shown in the exhibit. Users are reporting that the assistant sometimes refuses to answer questions that are clearly within the scope of the provided data. What is the most likely issue?
A.The system message encourages the model to err on the side of caution, leading to false-negative refusals.
B.The system message explicitly prohibits making up information, which is correct behavior.
C.The system message does not include instructions to use the provided data.
D.The temperature parameter is set too high, causing the model to hallucinate.
AnswerA
The instruction 'say I don't know' and not make up information can cause the model to refuse when uncertain.
Why this answer
The system message likely contains overly cautious language (e.g., 'only answer if you are certain' or 'do not speculate'), which causes the model to refuse answering even when the data clearly supports the response. This is a known behavior in Azure OpenAI chat completions where the system message's tone and constraints directly influence refusal rates, leading to false-negative refusals.
Exam trap
Microsoft often tests the misconception that refusal issues are caused by missing data instructions or high temperature, when in fact the root cause is the system message's overly cautious phrasing that induces false-negative refusals.
How to eliminate wrong answers
Option B is wrong because explicitly prohibiting making up information is a standard best practice to reduce hallucination, not a cause of false-negative refusals; it does not inherently make the model overly cautious. Option C is wrong because the system message in the exhibit (as described) does include instructions to use the provided data, so the issue is not a missing directive but the cautious phrasing. Option D is wrong because a high temperature parameter increases randomness and creativity, leading to hallucinations or off-topic responses, not systematic refusal to answer within scope.
A developer uses the Azure OpenAI API to generate code. They want to ensure that the generated code is in Python. Which parameter should they set?
A.temperature
B.max_tokens
C.system message
D.top_p
AnswerC
System message can guide the model to output Python code.
Why this answer
The system message parameter in the Azure OpenAI API is used to set the behavior and context of the assistant, including specifying the desired output format or language. By setting the system message to something like 'You are a helpful assistant that always generates Python code,' the developer can instruct the model to produce Python code consistently. This is the correct parameter for guiding the model's response style and content.
Exam trap
The trap here is that candidates confuse parameters that control randomness (temperature, top_p) or response length (max_tokens) with the system message, which is the correct mechanism for setting output language or format constraints.
How to eliminate wrong answers
Option A is wrong because temperature controls the randomness of the output, not the language or format of the generated code. Option B is wrong because max_tokens limits the length of the response, not the programming language. Option D is wrong because top_p (nucleus sampling) affects the diversity of token selection, not the language or content type.
Which TWO actions can you take to reduce the cost of using Azure OpenAI Service for a chat application?
Select 2 answers
A.Increase the frequency penalty.
B.Enable content filtering.
C.Set the max_tokens parameter to a lower value.
D.Increase the max_tokens parameter to allow longer responses.
E.Use a smaller model like GPT-3.5 instead of GPT-4.
AnswersC, E
Reduces token count per response.
Why this answer
Reducing the max_tokens parameter directly limits the number of tokens generated per API call, which lowers the cost since Azure OpenAI Service bills per token (both input and output). By capping the response length, you avoid paying for unnecessarily long completions.
Exam trap
The trap here is that candidates often confuse cost-saving techniques with performance-tuning parameters, mistakenly thinking that adjusting penalty settings or content filtering reduces token consumption, when in fact only limiting token output or using a cheaper model directly lowers the bill.
You are deploying an Azure AI Search solution that indexes customer support tickets. The solution must support fuzzy search for misspelled keywords and boost results from premium customers. Which two features should you configure?
A.Scoring profiles
B.Synonym maps
C.Suggesters
D.Custom analyzers
E.Fuzzy search parameters
AnswerA, E
Scoring profiles allow boosting based on fields like customer tier.
Why this answer
Scoring profiles allow you to boost search results based on specific criteria, such as customer tier, by assigning higher weights to certain fields or adding functions that increase scores for premium customers. This directly supports the requirement to prioritize results from premium customers.
Exam trap
The trap here is that candidates often confuse fuzzy search parameters with custom analyzers or synonym maps, but fuzzy search is a query-time parameter (e.g., searchMode=any&fuzzy=true) that does not require index-level configuration, while scoring profiles are the correct mechanism for boosting results.
How to eliminate wrong answers
Option B is wrong because synonym maps expand queries to include equivalent terms but do not handle misspellings or boost results by customer tier. Option C is wrong because suggesters enable autocomplete and search-as-you-type suggestions, not fuzzy matching or scoring boosts. Option D is wrong because custom analyzers define tokenization and text normalization rules but do not provide built-in fuzzy search parameters or scoring profile boosts.
You are building a customer support chatbot using Azure OpenAI Service. The chatbot must only answer questions related to the company's products and policies. It should refuse to answer off-topic questions. You need to implement this restriction effectively. What should you do?
A.Implement Retrieval Augmented Generation (RAG) with your product and policy data
B.Fine-tune the model on a dataset of product-related conversations
C.Use a system message that instructs the model to stay on topic
D.Set the temperature parameter to 0 to reduce randomness
AnswerA
RAG grounds the model in your data, ensuring answers are only from that data.
Why this answer
Retrieval Augmented Generation (RAG) grounds the model's responses in your specific product and policy data by retrieving relevant documents from a vector database (e.g., Azure Cognitive Search) and injecting them into the prompt. This ensures the chatbot can only answer questions that have matching context in your data, and it naturally refuses off-topic queries because no relevant documents are retrieved, allowing the system to return a default refusal message. RAG provides a dynamic, data-driven boundary without modifying the underlying model.
Exam trap
Microsoft often tests the misconception that a system message or fine-tuning alone can reliably enforce content restrictions, when in practice RAG provides a grounded, data-driven boundary that prevents off-topic responses by design.
How to eliminate wrong answers
Option B is wrong because fine-tuning adapts the model's behavior on a fixed dataset, but it does not prevent the model from generating off-topic answers; the model can still hallucinate or respond to unrelated queries outside the training distribution. Option C is wrong because a system message is a soft instruction that the model can ignore, especially for adversarial or ambiguous off-topic prompts, and it lacks a hard enforcement mechanism to block out-of-scope questions. Option D is wrong because setting the temperature parameter to 0 only reduces randomness in token selection, making outputs more deterministic, but it does not constrain the topic or domain of the response.
Match each Azure AI pricing tier to its description.
Drag a concept onto its matching description — or click a concept then click the description.
Concepts
Matches
Limited transactions per month for evaluation
Production tier with higher throughput
Higher throughput than S0
Even higher throughput for large workloads
Highest throughput tier
Why these pairings
Common Azure Cognitive Services pricing tiers include Free (limited free transactions), Standard S0 (pay-as-you-go moderate), and Premium (enterprise high-throughput).
Match each Azure Bot Service channel to its description.
Drag a concept onto its matching description — or click a concept then click the description.
Concepts
Matches
Custom client communication
Embedded chat in web pages
Integration with Teams
Integration with Slack
Integration with Facebook
Why these pairings
The correct matches are: Direct Line for custom client communication, Web Chat for embeddable chat, Microsoft Teams for team collaboration, and Facebook Messenger for Messenger integration. Common mistakes include confusing Direct Line and Web Chat definitions.
Your organization uses Azure AI Language to analyze customer support tickets. You need to ensure that personally identifiable information (PII) is detected and redacted before further processing. Which TWO features should you use?
Select 2 answers
A.Entity linking
B.Redaction
C.PII detection
D.Conversation summarization
E.Key phrase extraction
AnswersB, C
Redaction removes or masks the detected PII.
Why this answer
Redaction (Option B) is correct because it is the process of masking or removing detected PII from text, which directly meets the requirement to ensure PII is redacted before further processing. Azure AI Language provides a dedicated PII detection and redaction feature that identifies categories such as names, phone numbers, and credit card numbers, and then replaces them with placeholders like '********' or entity type labels.
Exam trap
This question tests the distinction between detection and redaction, where candidates mistakenly think that simply detecting PII (Option C) satisfies the requirement to 'ensure PII is redacted,' but the question explicitly requires both detection and redaction, making Option B necessary alongside Option C.
A multinational corporation uses Azure AI Language to analyze customer feedback in multiple languages. The solution must detect the language of incoming text and then perform sentiment analysis. Which approach minimizes latency and cost?
A.Use the sentiment analysis API with multilingual support
B.Use the Translator service to translate text to English, then call sentiment analysis
C.Store text in Azure AI Search and use cognitive skills for sentiment
D.Call the language detection API followed by the sentiment analysis API
AnswerD
This is the correct approach because you first call the language detection API to identify the language, then call the sentiment analysis API with that language code. This combination directly meets the requirement with minimal latency and cost, avoiding translation or indexing overhead.
Why this answer
The requirement is to detect the language of incoming text and then perform sentiment analysis. The sentiment analysis API in Azure AI Language does not detect language; it requires the language to be specified. Therefore, you must first call the language detection API to identify the language, then pass that language code to the sentiment analysis API.
While this involves two API calls, it is still more efficient than translating text (Option B) which adds significant latency and cost, or using cognitive skills with Azure AI Search (Option C) which adds indexing overhead. Option A is wrong because the sentiment analysis API cannot detect language on its own; it requires the language as an input parameter. Thus, Option D minimizes latency and cost by only performing the necessary steps without unnecessary transformations.
Exam trap
The trap is that candidates may assume the sentiment analysis API automatically detects language, but in reality it requires the language to be specified. Alternatively, they might think translation is needed, but language detection plus native sentiment analysis is more efficient for multilingual support.
How to eliminate wrong answers
Option B is wrong because translating text to English before sentiment analysis adds the latency and cost of an extra Translator API call, and translation may alter nuances or sentiment, reducing accuracy. Option C is wrong because storing text in Azure AI Search and using cognitive skills introduces unnecessary infrastructure overhead, indexing delays, and additional costs for search and skillset execution, which is not optimal for a simple real-time sentiment analysis pipeline. Option D is wrong because calling the language detection API first, then the sentiment analysis API, requires two separate API calls, doubling the latency and cost compared to using the single multilingual sentiment API that handles detection internally.
A healthcare organization deploys an Azure AI Language Service custom entity recognition model to extract medical conditions from clinical notes. During testing, the model fails to recognize rare diseases mentioned in the training data. What is the most likely cause?
A.The model requires more labeled examples of rare diseases
B.The entity length exceeds the maximum allowed for the model
C.The model is overfitting to common diseases
D.The training data is imbalanced with too many common diseases
AnswerA
Custom entity models need adequate examples per entity; rare diseases often lack sufficient labeled data.
Why this answer
The custom entity recognition model in Azure AI Language Service requires sufficient labeled examples for each entity type to learn patterns effectively. Rare diseases with few training examples will have lower recall because the model lacks the statistical evidence to generalize those entities. This is a classic data sparsity issue in supervised NLP models.
Exam trap
The trap here is that candidates confuse data imbalance (Option D) with insufficient labeled examples, but the core issue is that rare diseases lack enough training instances for the model to learn their contextual patterns, not just that common diseases dominate the dataset.
How to eliminate wrong answers
Option B is wrong because Azure AI Language Service custom NER does not impose a hard maximum entity length that would cause failure on rare diseases; the model can handle variable-length entities as long as they are properly labeled. Option C is wrong because overfitting to common diseases would cause the model to perform poorly on unseen data, not specifically on rare diseases that were present in training data. Option D is wrong because while data imbalance can reduce performance on minority classes, the most likely direct cause is insufficient labeled examples of rare diseases, not merely the presence of too many common diseases.
You have configured an Azure AI Search indexer with a Cosmos DB data source as shown in the exhibit. The indexer runs successfully, but you notice that the index is missing some documents that were recently added to Cosmos DB. What is the most likely cause?
A.The indexer is not configured to track changes using _ts.
B.The container name is misspelled.
C.The high water mark is not being updated correctly, causing some documents to be skipped.
D.The query does not select all fields required by the index.
AnswerC
If the high water mark is not updated, documents with _ts <= high water mark are skipped.
Why this answer
The most likely cause is that the high water mark (the _ts value) is not being updated correctly, causing the indexer to skip documents that were recently added. Azure AI Search uses a high water mark strategy to track changes from Cosmos DB, and if the _ts value is not properly recorded or the indexer's change tracking logic fails, new documents may be missed even though the indexer runs successfully.
Exam trap
The trap here is that candidates often assume the indexer must be failing entirely or that a configuration error is obvious, but the question describes a successful run with missing documents, which points to a subtle change tracking issue rather than a connectivity or schema mismatch.
How to eliminate wrong answers
Option A is wrong because the indexer is already configured to track changes using _ts (as shown in the exhibit), so this is not the cause. Option B is wrong because if the container name were misspelled, the indexer would fail to connect and would not run successfully. Option D is wrong because the query not selecting all fields required by the index would cause indexing errors or missing fields, not the omission of entire documents that were recently added.
You are using Azure OpenAI Service to generate code snippets. The output must be safe and free of security vulnerabilities. Which TWO practices should you follow? (Select TWO.)
Select 2 answers
A.Increase the temperature parameter to encourage diversity
B.Rely on the model's built-in safety features
C.Use Azure AI Content Safety to filter outputs
D.Fine-tune the model with a dataset of secure code examples
E.Include a system message instructing the model to follow secure coding practices
AnswersC, E
Content safety can detect and block harmful code.
Why this answer
Azure AI Content Safety is a dedicated service that provides an additional layer of filtering for harmful or inappropriate content, including security vulnerabilities, beyond what the model itself offers. It allows you to define custom severity thresholds and blocklists, ensuring that generated code snippets are safe before they reach the user. This is a recommended practice for production deployments to mitigate risks like injection attacks or exposure of sensitive patterns.
Exam trap
The trap here is that candidates often assume the model's built-in safety features are sufficient (Option B) or that increasing temperature (Option A) improves safety by adding randomness, when in fact Azure AI Content Safety is the explicit, exam-tested tool for output filtering in generative AI solutions.
Which THREE components are required to build a custom named entity recognition (NER) model in Azure Cognitive Service for Language?
Select 3 answers
A.A QnA Maker knowledge base for entity definitions.
B.A LUIS application to handle entity extraction.
C.A set of labeled documents for training.
D.A language service project with a custom NER schema.
E.A deployed endpoint for the custom NER model.
AnswersC, D, E
Labeled data is essential for training.
Why this answer
A custom NER model in Azure Cognitive Service for Language requires a set of labeled documents for training. These labeled documents define the entities and their spans within text, which the model uses to learn patterns for extraction. Without labeled data, the model cannot be trained to recognize custom entities.
Exam trap
The trap here is that candidates often confuse the separate Azure AI services (QnA Maker, LUIS, Language service) and assume they are interchangeable for custom NER, when in fact each has a distinct role and pipeline.
Which THREE components are required to build a custom chat application using Azure OpenAI Service that can answer questions based on your own private data?
Select 3 answers
A.The 'Add your data' feature configured in Azure OpenAI Studio.
B.Azure AI Search index.
C.An Azure OpenAI Service deployment.
D.A fine-tuned custom model.
E.An Azure OpenAI embeddings model deployment.
AnswersA, B, C
Enables grounding on private data.
Why this answer
The 'Add your data' feature in Azure OpenAI Studio provides a no-code interface to connect your private data sources (e.g., Azure Blob Storage, local files) to an Azure OpenAI chat model. It automatically chunks the data, creates an Azure AI Search index, and configures the retrieval-augmented generation (RAG) pipeline, enabling the model to answer questions grounded in your proprietary content without fine-tuning.
Exam trap
The trap here is that candidates often confuse fine-tuning (Option D) with retrieval-augmented generation, assuming that custom data requires model retraining, when in fact the 'Add your data' feature uses a RAG approach that does not modify the base model.
Drag a concept onto its matching description — or click a concept then click the description.
Concepts
Matches
Language Understanding Intelligent Service
Service to create a question and answer bot
Service to build custom image classifiers
Convert spoken language to text
Extract insights from text like key phrases
Why these pairings
Correct matches: Azure Cognitive Services provides pre-built AI APIs; Azure Machine Learning is for ML lifecycle; Azure Bot Service for building bots. Common confusion involves swapping definitions of Cognitive Services and Machine Learning.
A healthcare organization uses Azure AI Vision to analyze medical images for abnormalities. The solution must ensure patient data is not stored in the service's logs. What should you configure?
A.Use a private endpoint to access the Azure AI Vision resource.
B.Set the logging level to 'None' in the Azure AI Vision resource diagnostic settings.
C.Enable Customer-Managed Keys (CMK) for the resource.
D.Use Azure Policy to restrict data residency.
AnswerB
Disabling logging prevents data from being stored.
Why this answer
Setting the logging level to 'None' in the Azure AI Vision resource diagnostic settings prevents any patient data from being captured in the service's logs. This directly addresses the requirement to ensure patient data is not stored, as diagnostic logs can contain request and response payloads that include medical image metadata or analysis results.
Exam trap
The trap here is that candidates often confuse network security (private endpoints) or encryption (CMK) with data logging controls, failing to realize that logs are a separate data storage mechanism that must be explicitly disabled to prevent data retention.
How to eliminate wrong answers
Option A is wrong because using a private endpoint secures network traffic to the Azure AI Vision resource but does not control what data is stored in logs; it only prevents exposure over the public internet. Option C is wrong because Customer-Managed Keys (CMK) encrypt data at rest but do not prevent data from being written to logs; they only control encryption keys. Option D is wrong because Azure Policy restricts resource configurations and compliance, not the logging behavior of a specific resource; it cannot prevent data from being stored in logs once the resource is configured to log.
Your company wants to build a custom generative AI model that generates architectural designs. The model should be trained on the company's proprietary dataset of floor plans and designs. Which Azure service should you use?
A.Azure OpenAI Service
B.Azure Machine Learning
C.Azure AI Vision
D.Azure AI Document Intelligence
AnswerB
Azure Machine Learning supports training custom generative models using your own data.
Why this answer
Azure Machine Learning is the correct service because it provides a comprehensive platform for training custom generative AI models using your own proprietary datasets. It supports deep learning frameworks like PyTorch and TensorFlow, enabling you to build, train, and deploy a custom generative model for architectural designs, whereas Azure OpenAI Service is limited to pre-trained models and fine-tuning, not custom training from scratch.
Exam trap
The trap here is that candidates often confuse 'fine-tuning' (offered by Azure OpenAI Service) with 'custom training from scratch' (offered by Azure Machine Learning), leading them to incorrectly select Azure OpenAI Service when the question explicitly requires training a model on proprietary data with custom architecture.
How to eliminate wrong answers
Option A is wrong because Azure OpenAI Service only allows fine-tuning of existing pre-trained models (like GPT-4) on your data, not training a custom generative model from scratch on proprietary floor plans; it lacks the flexibility to define custom architectures. Option C is wrong because Azure AI Vision is designed for image analysis tasks (e.g., object detection, OCR) and does not support generative model training or generation of new designs. Option D is wrong because Azure AI Document Intelligence is specialized for extracting structured data from documents (e.g., forms, invoices) and cannot be used to train or generate architectural designs.
You have deployed a generative AI model using Azure Machine Learning. The model is used for generating financial reports. You need to monitor the model's performance and detect data drift in the input data. What should you use?
A.Azure Machine Learning data drift monitoring
B.Azure Monitor
C.Application Insights
D.Azure AI Language
AnswerA
This feature monitors changes in input data distribution compared to training data, alerting to drift.
Why this answer
Azure Machine Learning data drift monitoring is the correct choice because it is specifically designed to detect statistical changes in input data over time, which is critical for generative AI models used in financial reporting where data distributions can shift due to market changes or new regulations. It compares the current input data distribution against a baseline dataset using metrics like Wasserstein distance or Population Stability Index, and triggers alerts when drift exceeds a threshold. This ensures the model's outputs remain reliable and compliant, which is a core requirement for generative AI solutions in regulated industries.
Exam trap
The trap here is that candidates confuse general monitoring tools (Azure Monitor, Application Insights) with the specialized data drift detection capability in Azure Machine Learning, assuming any monitoring tool can detect statistical drift in input data.
How to eliminate wrong answers
Option B is wrong because Azure Monitor is a general-purpose observability service for infrastructure and application metrics (e.g., CPU, memory, request rates), not for detecting statistical data drift in model inputs. Option C is wrong because Application Insights focuses on application performance monitoring (APM) and telemetry (e.g., request failures, dependencies), not on comparing data distributions or detecting drift in input features. Option D is wrong because Azure AI Language is a service for natural language processing tasks (e.g., sentiment analysis, key phrase extraction), not for monitoring model performance or detecting data drift in input data.
You are troubleshooting an agentic solution where the agent is not returning responses within acceptable time limits. You suspect the agent is making too many sequential calls to external tools. Which strategy should you recommend to reduce latency?
A.Increase the max token limit
B.Enable parallel tool execution
C.Add more tools to distribute the load
D.Reduce the thread history length
AnswerB
Parallel execution allows the agent to invoke multiple tools simultaneously.
Why this answer
Enabling parallel tool execution allows the agent to invoke multiple external tools simultaneously rather than sequentially, directly reducing the total latency caused by serial tool calls. This is a core optimization in agentic frameworks like Semantic Kernel or AutoGen, where tool calls are independent and can be dispatched concurrently.
Exam trap
The trap here is that candidates confuse throughput improvements (like adding tools or increasing token limits) with latency reduction, when the real bottleneck is the sequential dependency of tool calls.
How to eliminate wrong answers
Option A is wrong because increasing the max token limit does not affect the number or sequence of tool calls; it only allows longer responses, which can actually increase latency. Option C is wrong because adding more tools increases the workload and potential sequential calls, worsening latency rather than distributing load in a meaningful way. Option D is wrong because reducing thread history length may free context window space but does not change the sequential execution pattern of tool calls, so it has no direct impact on latency from tool orchestration.
A healthcare organization uses Azure AI Language to extract medical entities from clinical notes. The solution must comply with HIPAA. What should you configure?
A.Use Azure AI Language with HIPAA-enabled service and enable encryption
B.Configure data residency to a specific region
C.Enable content filtering
D.Isolate the service in a virtual network
AnswerA
Azure AI Language offers HIPAA-compliant tiers with data encryption.
Why this answer
Azure AI Language can be configured with a HIPAA-eligible service tier, and enabling encryption at rest and in transit ensures compliance with HIPAA's security and privacy rules. This configuration meets the healthcare organization's requirement to protect protected health information (PHI) when extracting medical entities from clinical notes.
Exam trap
The trap here is that candidates often confuse network isolation (VNet) or data residency with HIPAA compliance, not realizing that HIPAA mandates specific encryption and a signed BAA, which are only available with a HIPAA-eligible service tier.
How to eliminate wrong answers
Option B is wrong because configuring data residency to a specific region addresses data sovereignty but does not by itself ensure HIPAA compliance, which requires additional safeguards like encryption and access controls. Option C is wrong because content filtering in Azure AI Language is designed to block offensive or harmful content, not to enforce HIPAA compliance or protect PHI. Option D is wrong because isolating the service in a virtual network (VNet) provides network-level security but does not automatically meet HIPAA requirements; encryption and a HIPAA-eligible service tier are still necessary.
Your company uses Azure AI Content Safety to moderate user-generated content in a chat application. You need to detect and block sexual content in multiple languages. Which pre-built category should you configure?
A.Sexual
B.Self-harm
C.Hate
D.Violence
AnswerA
This category specifically detects sexual content.
Why this answer
Azure AI Content Safety provides pre-built severity-based categories for content moderation. The 'Sexual' category is specifically designed to detect and block explicit sexual content, including text and images, across multiple languages. This makes it the correct choice for your requirement to moderate sexual content in a chat application.
Exam trap
The trap here is that candidates may confuse 'Sexual' with broader categories like 'Hate' or 'Violence', not realizing that Azure AI Content Safety has a dedicated pre-built category for sexual content with specific detection capabilities.
How to eliminate wrong answers
Option B (Self-harm) is wrong because it focuses on content related to self-injury or suicide, not sexual material. Option C (Hate) is wrong because it targets hate speech based on protected attributes like race or religion, not sexual content. Option D (Violence) is wrong because it detects violent acts or threats, which are distinct from sexual content.
You need to generate realistic synthetic data using Azure OpenAI Service to train a machine learning model. The data must be diverse and cover edge cases. Which approach should you use?
A.Use prompt engineering with detailed instructions to generate varied examples.
B.Fine-tune the model on a small dataset of real examples.
C.Use Azure OpenAI embeddings to generate similar data points.
D.Set a high temperature parameter only.
AnswerA
Prompt engineering effectively controls output diversity and coverage.
Why this answer
Prompt engineering with detailed instructions allows you to explicitly control the diversity, structure, and edge-case coverage of generated synthetic data without requiring a pre-existing dataset. By crafting system messages and user prompts that specify variations in attributes, formats, and boundary conditions, you can produce a wide range of realistic examples that mimic real-world distributions, which is essential for training robust machine learning models.
Exam trap
The trap here is that candidates often assume fine-tuning or embeddings are the only ways to generate realistic data, overlooking that prompt engineering with detailed instructions is the most direct and flexible method for producing diverse synthetic data without requiring a pre-existing labeled dataset.
How to eliminate wrong answers
Option B is wrong because fine-tuning on a small dataset of real examples would bias the model toward the limited patterns in that dataset, reducing diversity and failing to generate edge cases, which contradicts the requirement for varied synthetic data. Option C is wrong because Azure OpenAI embeddings are used to measure semantic similarity between text inputs, not to generate new data points; they can retrieve or cluster existing data but cannot produce novel synthetic examples. Option D is wrong because setting a high temperature parameter alone increases randomness in token selection but does not provide the structured control needed to ensure coverage of specific edge cases or diverse scenarios; it may produce incoherent or irrelevant outputs without detailed prompt guidance.
You are using Azure AI Search to index a set of PDF documents. The index includes a 'content' field with the extracted text. Users report that when they search for 'budget forecast', documents containing only 'budget' or 'forecast' are ranked lower than expected. Which configuration change would improve the ranking for multi-word queries?
A.Add a separate field for each word in the document
B.Change the analyzer to a custom analyzer that splits on spaces only
C.Enable semantic search on the index
D.Set the 'content' field to a higher boosting value
AnswerC
Semantic search uses advanced ranking models that consider the meaning and relationship between words.
Why this answer
Semantic search in Azure AI Search uses deep neural network models to understand the intent and context of multi-word queries like 'budget forecast', rather than relying solely on keyword matching. This improves ranking by capturing the semantic relationship between terms, so documents containing both words (or conceptually related content) are ranked higher even if the exact phrase is not present.
Exam trap
The trap here is that candidates often confuse boosting (Option D) with semantic understanding, but boosting only increases term frequency weight and does not address the semantic gap between query terms in multi-word searches.
How to eliminate wrong answers
Option A is wrong because adding a separate field for each word would fragment the index and break the ability to score multi-word queries as a unit, degrading relevance rather than improving it. Option B is wrong because splitting on spaces only is the default behavior for standard analyzers; a custom analyzer that does the same would not change ranking behavior for multi-word queries. Option D is wrong because boosting the 'content' field increases the weight of all terms in that field uniformly, which does not specifically address the need to rank documents containing both query terms higher than those containing only one.
You are deploying a solution that uses Azure OpenAI Service to generate financial reports. You need to ensure the outputs are accurate and consistent. Which TWO parameters should you adjust? (Choose two.)
Select 2 answers
A.Set presence_penalty to 0.5.
B.Set temperature to 0.
C.Set max_tokens to 2000.
D.Set frequency_penalty to 0.7.
E.Set top_p to 0.1.
AnswersB, E
Low temperature makes output more deterministic.
Why this answer
Setting temperature to 0 makes the model deterministic, always choosing the most likely next token. This is critical for financial reports where consistency and reproducibility are required, as it eliminates randomness in the output.
Exam trap
The trap here is that candidates often confuse parameters that control creativity (temperature, top_p) with those that control repetition (frequency_penalty, presence_penalty) or output length (max_tokens), leading them to select options that do not directly address accuracy and consistency.
A financial services firm uses Azure AI Language to analyze earnings call transcripts. They need to extract key phrases and identify sentiment for each speaker's turn. Which approach should they use?
A.Call the prebuilt sentiment analysis API on the entire transcript
B.Split the transcript by speaker turns and call key phrase extraction and sentiment analysis on each part
C.Use QnA Maker to extract Q&A pairs per speaker
D.Use Text Analytics for health to extract entities and sentiment
AnswerB
Splitting by speaker turns allows per-speaker analysis.
Why this answer
The requirement is to extract key phrases and identify sentiment per speaker turn. The Azure AI Language key phrase extraction and sentiment analysis APIs operate on individual text inputs. By splitting the transcript by speaker turns, each segment can be analyzed independently, providing per-speaker insights.
Processing the entire transcript as a single document would aggregate sentiment and key phrases, losing per-speaker granularity.
Exam trap
The trap here is that candidates may assume the prebuilt sentiment analysis API can handle multi-speaker transcripts by default, but it processes the entire input as one document, so splitting by speaker turns is necessary for per-speaker granularity.
How to eliminate wrong answers
Option A is wrong because calling the prebuilt sentiment analysis API on the entire transcript would return a single overall sentiment score and key phrases for the whole document, not per-speaker turn, failing to meet the requirement for speaker-level analysis. Option C is wrong because QnA Maker (now part of Azure AI Language as custom question answering) is designed to extract question-answer pairs from FAQ-like content, not to perform key phrase extraction or sentiment analysis on conversational transcripts. Option D is wrong because Text Analytics for health is a specialized domain model for extracting medical entities (e.g., diagnoses, medications) and sentiment from clinical notes, not suitable for financial earnings call transcripts.
Refer to the exhibit. You are reviewing a content safety policy for an Azure AI Foundry deployment. The policy rate limits to 20 requests per minute. A user submits 50 requests in one minute. How many requests are allowed?
A.20
B.50
C.100
D.None, all are blocked.
AnswerA
The rate limit policy sets 20 requests per minute.
Why this answer
The content safety policy enforces a rate limit of 20 requests per minute. When a user submits 50 requests in one minute, the rate limiter allows only the first 20 requests and blocks the remaining 30. This is a standard token-bucket or sliding-window rate-limiting behavior in Azure AI Foundry, where exceeding the limit results in HTTP 429 (Too Many Requests) for excess requests.
Exam trap
Microsoft often tests the misconception that exceeding a rate limit blocks all requests, when in fact the limit is a threshold that allows the first N requests and denies the rest.
How to eliminate wrong answers
Option B is wrong because it assumes all 50 requests are allowed, ignoring the explicit rate limit of 20 per minute. Option C is wrong because 100 is not related to any limit in the policy; it may confuse the rate limit with a burst or quota value. Option D is wrong because the policy does not block all requests; it allows up to the limit (20) and then blocks the excess.
You are using Azure AI Search to index customer support tickets. You want to automatically extract the customer's sentiment and key phrases from each ticket. Which Azure AI service should you integrate as a skillset?
A.Azure AI Document Intelligence
B.Azure AI Computer Vision
C.Azure AI Translator
D.Azure AI Language
AnswerD
Offers sentiment and key phrase extraction skills.
Why this answer
Azure AI Language provides pre-built capabilities for sentiment analysis and key phrase extraction, which are exactly the skills needed to process customer support ticket text. When integrated as a skillset in Azure AI Search, it enriches the indexing pipeline by automatically extracting these insights from each document. The other services focus on different modalities (vision, translation, document structure) and do not offer native sentiment or key phrase extraction.
Exam trap
The AI-102 exam often tests the distinction between Azure AI Language (for text analytics) and Azure AI Document Intelligence (for document structure extraction), leading candidates to mistakenly choose Document Intelligence when the task involves analyzing text content rather than extracting form fields.
How to eliminate wrong answers
Option A is wrong because Azure AI Document Intelligence is designed for extracting structured data (like tables, forms, and key-value pairs) from scanned documents, not for analyzing sentiment or extracting key phrases from text. Option B is wrong because Azure AI Computer Vision analyzes images and video, not text content, so it cannot perform sentiment analysis or key phrase extraction on ticket text. Option C is wrong because Azure AI Translator focuses on language translation, not on extracting sentiment or key phrases from the original language text.
A company is building a chatbot that must handle user queries in multiple languages. The chatbot uses Azure AI Language Service. Which feature should be used to detect the language of incoming messages before routing them to the appropriate language model?
A.Sentiment Analysis
B.Key Phrase Extraction
C.Language Detection
D.Entity Recognition
AnswerC
Language Detection identifies the language of text input.
Why this answer
Language Detection is the correct feature because it is specifically designed to identify the language of text input, returning a language name and a confidence score. In a multi-language chatbot, this detection step is essential to route the query to the appropriate language-specific model or handler. Azure AI Language Service provides a dedicated pre-built capability for language detection, which can be called via the REST API or SDK.
Exam trap
The trap here is that candidates confuse Language Detection with other text analytics features like Sentiment Analysis or Entity Recognition, assuming any 'analysis' feature can identify language, but only Language Detection is purpose-built for this task.
How to eliminate wrong answers
Option A is wrong because Sentiment Analysis determines the emotional tone (positive, negative, neutral) of text, not the language. Option B is wrong because Key Phrase Extraction identifies important words or phrases in the text, but does not detect the language. Option D is wrong because Entity Recognition identifies named entities like people, places, or organizations, not the language of the text.
A company uses Azure Face API to verify employee identities for building access. They need to ensure that only live faces are used, not photos or videos. Which feature should they enable?
A.Set a high confidence threshold for face matching.
B.Face identification with a large person group.
C.Enable liveness detection using session-based verification.
D.Face detection with attributes such as age and emotion.
AnswerC
Liveness detection checks for spoofing attacks.
Why this answer
Azure Face API's liveness detection with session-based verification is specifically designed to prevent spoofing attacks using photos, videos, or masks. It analyzes subtle cues such as micro-movements, texture, and depth to confirm the presence of a live person, ensuring that only live faces are accepted for identity verification.
Exam trap
The trap here is that candidates may confuse confidence thresholds or face attributes with liveness detection, not realizing that only session-based verification actively checks for spoofing through motion and depth analysis.
How to eliminate wrong answers
Option A is wrong because setting a high confidence threshold only increases the strictness of face matching scores, but does not differentiate between a live face and a spoofed image or video. Option B is wrong because face identification with a large person group is used to match a detected face against a database of enrolled persons, but it does not verify liveness or detect presentation attacks. Option D is wrong because face detection with attributes like age and emotion extracts demographic and emotional information from a face, but it cannot determine whether the face is live or a reproduction.
An organization uses Microsoft Copilot Studio to create an agent for IT support. The agent should be able to reset passwords, unlock accounts, and look up user information by connecting to on-premises Active Directory via Microsoft Entra ID. Which type of authentication should be used for the agent to access these actions?
A.Basic authentication with username and password.
B.API key authentication to the on-premises API.
C.Certificate-based authentication for the agent's identity.
D.OAuth 2.0 authentication with Microsoft Entra ID using on-behalf-of flow.
AnswerD
On-behalf-of flow allows the agent to obtain tokens to call downstream APIs on behalf of the user.
Why this answer
The agent needs to securely access on-premises Active Directory through Microsoft Entra ID. OAuth 2.0 with the on-behalf-of (OBO) flow allows Copilot Studio to obtain a token for the agent's identity, then exchange it for a token to call downstream APIs (like Microsoft Graph or a custom API) that interact with AD. This provides delegated access and supports modern authentication without exposing static credentials.
Exam trap
The trap here is that candidates confuse certificate-based authentication (often used for daemon apps) with the on-behalf-of flow, not realizing that Copilot Studio agents require delegated user context and token exchange rather than a static identity.
How to eliminate wrong answers
Option A is wrong because Basic authentication sends credentials in plaintext (base64-encoded) and is not supported for modern cloud-to-on-premises integrations via Entra ID; it lacks token-based security and is deprecated for Microsoft Graph. Option B is wrong because API key authentication is typically used for stateless API access and does not integrate with Entra ID's identity framework; it cannot handle delegated user context or on-behalf-of flows required for the agent's actions. Option C is wrong because certificate-based authentication is used for service-to-service or application-only scenarios (e.g., daemon apps) and does not support the on-behalf-of flow needed to act on behalf of a user; it also requires complex certificate management and is not the standard for Copilot Studio agents.
You are a generative AI engineer at a financial services company. The company uses Azure OpenAI Service to generate investment summaries. You have deployed a GPT-4 model with a content filter set to 'Low' for hate speech. The model frequently generates summaries that include biased language against certain demographics. You need to reduce biased outputs while maintaining the ability to generate detailed financial analysis. You cannot afford to retrain the model. You have the following options: A) Change the content filter severity to 'High' for all categories, B) Add a system message instructing the model to avoid bias and provide examples of unbiased summaries in the prompt, C) Use the Azure AI Language service to detect bias in the output and regenerate if bias is found, D) Deploy a different model like GPT-3.5 which has less bias. Which course of action should you take?
A.Use the Azure AI Language service to detect bias in the output and regenerate if bias is found.
B.Add a system message instructing the model to avoid bias and provide examples of unbiased summaries in the prompt.
C.Change the content filter severity to 'High' for all categories.
D.Deploy a different model like GPT-3.5 which has less bias.
AnswerB
This approach guides the model to produce unbiased outputs by providing explicit instructions and examples.
Why this answer
Adding a system message and examples in the prompt (few-shot learning) can effectively reduce biased outputs without retraining. Option A is wrong because increasing content filter severity may block legitimate financial analysis content. Option C is wrong because detecting bias after generation and regenerating is inefficient and may still produce biased output.
Option D is wrong because GPT-3.5 may also exhibit bias and may not provide the same quality of financial analysis.