View Full Microsoft AI-901 Exam Dumps and Practice Test Dumps

 

Question 381. Which Azure service is designed to store and search enterprise documents for AI applications?

1. Azure AI Search
2. Azure DNS
3. Azure Functions
4. Azure Monitor

Correct Answer: 1. Azure AI Search

Explanation:

Azure AI Search is a cloud search service designed to index and retrieve information from structured and unstructured data. It can support keyword, semantic, vector, and hybrid search scenarios, making it useful for AI applications that need to locate relevant enterprise information. For example, a company can index policies, manuals, reports, or product documentation and then retrieve relevant content when a user asks a question. Azure AI Search is commonly used with retrieval-augmented generation applications because it can provide relevant documents or passages that are then supplied as context to a generative AI model.

Question 382. Which technique allows a generative AI model to use retrieved documents as context when answering a question?

1. Image segmentation
2. Retrieval-augmented generation
3. Regression
4. Clustering

Correct Answer: 2. Retrieval-augmented generation

Explanation:

Retrieval-augmented generation, or RAG, combines information retrieval with generative AI. When a user submits a question, the application first searches a connected knowledge source for relevant information. The retrieved content is then provided to the generative model as context, allowing the model to formulate an answer based on that information. RAG is useful when an application needs access to private, current, or organization-specific information that may not have been part of the model’s original training data. Good retrieval, authorization, grounding, and evaluation are important components of a reliable RAG solution.

Question 383. What is the purpose of a system message in a generative AI application?

1. To define instructions and behavioral guidance for the model
2. To store database backups
3. To calculate classification accuracy
4. To create vector indexes

Correct Answer: 1. To define instructions and behavioral guidance for the model

Explanation:

A system message provides high-level instructions that guide how a generative AI model should behave during an interaction. It can establish the assistant’s role, response style, constraints, priorities, and other behavioral requirements. For example, an application may instruct the model to act as a customer-support assistant and answer using information supplied from an approved knowledge base. System instructions are different from the user’s individual question because they establish broader behavior for the interaction. Developers should still evaluate system instructions carefully because clear prompts alone cannot guarantee correct, safe, or reliable model behavior.

Question 384. Which prompting technique provides examples of desired input and output behavior to guide the model?

1. Zero-shot prompting
2. Data normalization
3. Few-shot prompting
4. Feature scaling

Correct Answer: 3. Few-shot prompting

Explanation:

Few-shot prompting provides the model with a small number of examples that demonstrate how a task should be performed. These examples can show the expected format, classification categories, writing style, or reasoning pattern. The model can then use the examples as guidance when processing a new input. Zero-shot prompting, by contrast, provides instructions without task-specific examples. Few-shot prompting can be useful when a task has a particular expected structure that may not be fully communicated through instructions alone. The examples should be representative, accurate, and concise so they do not introduce confusing or misleading patterns.

Question 385. What is the primary purpose of tokenization in a language model?

1. To convert text into smaller units that the model can process
2. To encrypt confidential documents
3. To create database tables
4. To monitor network traffic

Correct Answer: 1. To convert text into smaller units that the model can process

Explanation:

Tokenization converts text into smaller units called tokens that a language model can process. Depending on the tokenizer and language, a token may represent a complete word, part of a word, punctuation, or another text unit. Tokenization is important because models process token sequences rather than raw human-readable text. The number of tokens can also affect context-window usage, latency, and cost in applications that use language models. Developers should therefore consider token limits when designing prompts and retrieved context, especially when applications process long documents or conversations.

Question 386. What does a model’s context window determine?

1. The maximum number of network connections
2. The amount of text or tokens the model can consider within a request
3. The number of database tables available
4. The size of an image file

Correct Answer: 2. The amount of text or tokens the model can consider within a request

Explanation:

A model’s context window defines how much tokenized information can be included in the model’s working context for a request. This can include system instructions, user prompts, conversation history, retrieved documents, and other input information. If an application supplies too much information, it may exceed the model’s supported context limit. Developers therefore need to manage prompts, conversation history, and retrieved content efficiently. Techniques such as summarization, relevant retrieval, and appropriate chunking can help keep the context focused while providing the model with the information it needs to produce useful responses.

Question 387. Which practice can help defend a generative AI application against prompt injection?

1. Removing all authentication
2. Treating retrieved content as trusted instructions
3. Separating instructions from untrusted retrieved content and applying appropriate controls
4. Disabling monitoring

Correct Answer: 3. Separating instructions from untrusted retrieved content and applying appropriate controls

Explanation:

Prompt injection occurs when malicious or unintended instructions are introduced into content that an AI application processes, potentially influencing the model’s behavior. A defense strategy should treat external or retrieved content as untrusted data rather than automatically treating it as authoritative instructions. Applications can use clear instruction boundaries, input validation, content filtering, authorization controls, tool restrictions, and monitoring to reduce the risk. No single technique guarantees complete protection. Developers should test applications with adversarial inputs and continuously evaluate how the system behaves when users or retrieved documents contain conflicting or malicious instructions.

Question 388. Which capability helps an AI application determine whether generated responses contain harmful categories of content?

1. Azure AI Content Safety
2. Azure Blob Storage
3. Azure Virtual Network
4. Azure Machine Learning compute

Correct Answer: 1. Azure AI Content Safety

Explanation:

Azure AI Content Safety provides capabilities for detecting and managing potentially harmful content in AI applications. Depending on the supported scenario, applications can analyze text or images for categories of unsafe material and use the results to apply appropriate controls. This can help developers create safer user experiences, particularly when applications accept user-generated content or produce generated responses. Content safety should be integrated into a broader responsible AI strategy that can include access controls, application policies, human review, monitoring, testing, and incident response. Automated content detection should not be treated as a complete replacement for governance and human oversight.

Question 389. Which responsible AI principle focuses on ensuring an AI system performs appropriately across different relevant groups?

1. Fairness
2. Compression
3. Tokenization
4. Serialization

Correct Answer: 1. Fairness

Explanation:

Fairness involves considering whether an AI system provides appropriate outcomes and performance across relevant groups and circumstances. A model can appear accurate overall while performing differently for particular groups because of differences in training data, representation, features, or other factors. Evaluating performance across relevant groups can help identify disparities that an overall metric might hide. Developers should use representative evaluation data and consider appropriate fairness measures for the specific scenario. Fairness does not mean that every application must use identical metrics or outcomes; the appropriate evaluation depends on the system’s intended use and potential impacts.

Question 390. Which responsible AI principle emphasizes making an AI system’s behavior understandable to relevant stakeholders?

1. Storage
2. Transparency
3. Compression
4. Randomization

Correct Answer: 2. Transparency

Explanation:

Transparency focuses on providing understandable information about an AI system, including how it is designed, what it is intended to do, what data or processes may influence its behavior, and what limitations should be considered. Appropriate transparency helps users and other stakeholders understand when they are interacting with AI and how its outputs should be interpreted. The amount and type of information required depend on the application and its risks. Documentation, clear user communication, model information, and explanations of limitations can all contribute to transparency without suggesting that every model can provide a complete explanation of its internal reasoning.

Question 391. Which responsible AI principle requires organizations to consider how an AI system can be made dependable and resistant to failures?

1. Reliability and safety
2. Vectorization
3. Data compression
4. Tokenization

Correct Answer: 1. Reliability and safety

Explanation:

Reliability and safety involve designing AI systems to operate consistently and safely under expected conditions while reducing the risk of harmful failures. This includes testing the system with representative scenarios, identifying edge cases, monitoring production behavior, and establishing safeguards for known risks. Developers should consider how the application behaves when inputs are incomplete, unexpected, adversarial, or outside the system’s intended scope. Monitoring after deployment is also important because real-world conditions can change. Reliability and safety are therefore ongoing responsibilities rather than properties that can be established only during initial development.

Question 392. Which practice can help reduce privacy risks when preparing data for an AI application?

1. Collecting unnecessary personal information
2. Applying data minimization and appropriate access controls
3. Sharing all records publicly
4. Removing authentication

Correct Answer: 2. Applying data minimization and appropriate access controls

Explanation:

Data minimization reduces privacy exposure by limiting collection and processing to information that is necessary for the intended purpose. Access controls further reduce risk by ensuring that only authorized people, services, or applications can access sensitive information. Other privacy practices can include appropriate retention policies, encryption, de-identification where suitable, secure data handling, and careful logging. These controls should be considered throughout the AI lifecycle rather than only after a system has been deployed. Protecting data is particularly important when AI applications process personal, confidential, financial, or organizational information.

Question 393. Which Azure service can extract structured information from invoices and other documents?

1. Azure AI Document Intelligence
2. Azure DNS
3. Azure Firewall
4. Azure Virtual Network

Correct Answer: 1. Azure AI Document Intelligence

Explanation:

Azure AI Document Intelligence provides capabilities for extracting information from documents such as invoices, receipts, forms, and other business records. It can analyze document layout and identify relevant fields, text, tables, and other structured information depending on the model and scenario. This can reduce the amount of manual data entry required when processing large numbers of documents. Document Intelligence can support both prebuilt models and customized document-processing scenarios. The extracted information should still be validated because document quality, layout variations, handwriting, scanning issues, and unusual formats can affect automated extraction accuracy.

Question 394. Which Azure AI capability converts spoken audio into written text?

1. Text-to-speech
2. Speech-to-text
3. Image classification
4. OCR

Correct Answer: 2. Speech-to-text

Explanation:

Speech-to-text converts spoken audio into written text. It is commonly used for applications such as transcription, voice-controlled interfaces, meeting summaries, call analysis, and accessibility features. Azure AI Speech provides speech recognition capabilities that can process supported audio and produce text output. Performance can depend on factors such as audio quality, background noise, speaker characteristics, language, pronunciation, and application configuration. Developers should evaluate speech recognition using realistic recordings and conditions rather than relying only on ideal test audio. For applications involving sensitive conversations, privacy and data-handling requirements should also be considered.

Question 395. Which Azure AI capability converts written text into spoken audio?

1. Speech-to-text
2. Named entity recognition
3. Text-to-speech
4. Key phrase extraction

Correct Answer: 3. Text-to-speech

Explanation:

Text-to-speech converts written text into synthesized spoken audio. Azure AI Speech provides text-to-speech capabilities that can be used in applications such as virtual assistants, accessibility tools, interactive learning systems, announcements, and voice interfaces. Developers can select supported voices and configure appropriate language or speaking characteristics for the intended scenario. The quality of generated speech can depend on the selected voice, language, pronunciation, and input text. Applications should also consider user expectations and responsible use when synthetic voices are used, particularly when the generated voice could be confused with a real person.

Question 396. Which Azure AI Language capability identifies important words or phrases in a document?

1. Key phrase extraction
2. Image segmentation
3. Object detection
4. Speech synthesis

Correct Answer: 1. Key phrase extraction

Explanation:

Key phrase extraction identifies important concepts or phrases within text. It can help applications summarize the main topics represented in documents without requiring a full generative summary. For example, an organization could use key phrase extraction to identify important subjects from customer feedback, support requests, or business reports. This capability is different from named entity recognition, which focuses on identifying entities such as people, organizations, locations, or dates. Key phrase extraction can be useful for indexing, categorization, analytics, and search scenarios where identifying the main concepts in large volumes of text is valuable.

Question 397. Which Azure AI Language capability identifies people, organizations, locations, and other entities in text?

1. Sentiment analysis
2. Named entity recognition
3. Text summarization
4. Language detection

Correct Answer: 2. Named entity recognition

Explanation:

Named entity recognition, or NER, identifies entities within text and can categorize them according to supported entity types. Examples can include people, organizations, locations, dates, quantities, and other recognized categories. NER can help applications understand important references in documents and conversations. For example, a business application might extract organization names and locations from customer messages for indexing or analysis. NER differs from key phrase extraction because NER focuses on recognized entity types, while key phrase extraction identifies important concepts or phrases more generally. Both capabilities can be useful for processing large collections of unstructured text.

Question 398. Which machine learning approach is most appropriate when training data contains labeled examples of the desired output?

1. Unsupervised learning
2. Reinforcement learning
3. Supervised learning
4. Random search

Correct Answer: 3. Supervised learning

Explanation:

Supervised learning uses labeled training examples where the desired output is known. The model learns a relationship between input features and corresponding labels or target values. Common supervised learning tasks include classification, where the target is a category, and regression, where the target is a numerical value. For example, a model trained with historical customer records labeled as retained or canceled could learn to predict customer retention categories. The quality and representativeness of the labeled data are important because the model can learn inaccurate or biased patterns when the training examples are incomplete, inconsistent, or poorly representative.

Question 399. Which machine learning approach groups similar data points without predefined labels?

1. Supervised learning
2. Regression
3. Classification
4. Unsupervised learning

Correct Answer: 4. Unsupervised learning

Explanation:

Unsupervised learning works with data where predefined target labels are not supplied. One common unsupervised technique is clustering, which groups data points according to similarities in their features. For example, a business could use clustering to identify customer segments based on purchasing behavior without first defining the customer groups. The resulting clusters still require interpretation because the algorithm does not automatically determine the business meaning of each group. Developers should evaluate whether the discovered patterns are useful and meaningful for the intended application rather than assuming that every mathematically identified cluster represents a meaningful real-world category.

Question 400. What is an important final step before releasing an AI application to production?

1. Remove all monitoring
2. Evaluate the application against intended requirements, risks, and representative scenarios
3. Disable security controls
4. Delete the validation data

Correct Answer: 2. Evaluate the application against intended requirements, risks, and representative scenarios

Explanation:

Before releasing an AI application to production, teams should evaluate whether it meets its intended requirements and behaves appropriately under realistic conditions. Evaluation can include functional performance, accuracy or quality metrics, safety testing, security and privacy checks, fairness considerations, edge-case testing, latency, and other requirements relevant to the application. Representative scenarios are important because performance on a narrow test set may not reflect real-world usage. Organizations should also establish monitoring and response procedures before deployment so problems can be detected and addressed after release. Responsible AI development continues throughout the system’s operational lifecycle.