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

Question 261. Which Azure AI Foundry capability helps you compare available generative AI models before selecting one?

1. Model catalog
2. Azure Monitor
3. Azure Storage Explorer
4. Azure DNS

Correct Answer: 1. Model catalog

Explanation:

The Azure AI Foundry model catalog provides a centralized place to explore and compare available models for different generative AI scenarios. Developers can review model capabilities, supported tasks, deployment options, and other characteristics before choosing a model for an application. This is useful because different models can vary significantly in language understanding, reasoning, speed, cost, context capacity, and output quality. Selecting a model should therefore be based on the application’s requirements rather than assuming that one model is appropriate for every workload. The model catalog helps teams evaluate available choices and identify a model that aligns with their specific use case.

Question 262. What is a key limitation of generative AI models that developers should consider when designing an application?

1. They can only process numerical data
2. They may generate plausible but incorrect information
3. They cannot generate natural-language responses
4. They always produce identical responses

Correct Answer: 2. They may generate plausible but incorrect information

Explanation:

Generative AI models can produce fluent and convincing responses that are not necessarily factually correct. This behavior is often associated with hallucination, where the model generates information that appears reasonable but is unsupported, inaccurate, or fabricated. Developers should therefore evaluate outputs and apply appropriate safeguards, especially in applications where accuracy is important. Grounding responses with trusted information, using retrieval systems, validating outputs, and maintaining human oversight can reduce the impact of incorrect responses. Understanding this limitation is important when designing responsible generative AI solutions because language quality alone does not guarantee factual accuracy.

Question 263. What does the term token generally represent in a generative AI model?

1. A physical server
2. A database table
3. A unit of text processed by the model
4. A security certificate

Correct Answer: 3. A unit of text processed by the model

Explanation:

A token is a unit of text that a generative AI model processes when interpreting an input or producing an output. A token may correspond to an entire word, part of a word, punctuation, or another small piece of text depending on the model’s tokenizer. Token usage matters because models have limits on how much text they can process within a context window. It can also affect usage costs for services that charge based on tokens. Understanding tokens helps developers design prompts efficiently, manage context length, and estimate the resources required by generative AI applications.

Question 264. Why is a model’s context window important when designing a generative AI application?

1. It determines the maximum amount of information the model can consider in a request
2. It controls the physical size of the server
3. It determines the network bandwidth
4. It replaces the need for prompts

Correct Answer: 1. It determines the maximum amount of information the model can consider in a request

Explanation:

A model’s context window defines how much input and conversational information the model can process as context for generating a response. This can include the user’s prompt, previous conversation content, retrieved documents, instructions, and other information supplied to the model. If an application sends more information than the model can accommodate, some information may need to be removed, summarized, or handled through another strategy. Context limits are therefore important when building chat applications, retrieval-augmented generation systems, and document-processing solutions. Developers need to manage context carefully so relevant information remains available to the model.

Question 265. Which technique can help reduce the risk of prompt injection in a generative AI application?

1. Treat every retrieved document as a trusted instruction
2. Disable all user input
3. Clearly separate instructions from untrusted external content
4. Remove authentication

Correct Answer: 3. Clearly separate instructions from untrusted external content

Explanation:

Prompt injection occurs when untrusted input attempts to influence a model’s behavior by inserting instructions that conflict with the application’s intended instructions. A useful defensive approach is to clearly distinguish trusted system instructions from user-provided or retrieved content. Applications should also validate inputs, restrict sensitive operations, avoid treating retrieved text as authoritative instructions, and apply appropriate access controls. For higher-risk scenarios, developers should add monitoring, output validation, and human approval for important actions. No single technique eliminates every prompt-injection risk, so applications should use multiple layers of protection when models interact with external or user-controlled content.

Question 266. Which approach provides human oversight for an AI system making important decisions?

1. Removing all review processes
2. Allowing the model to make every decision automatically
3. Increasing the number of tokens
4. Requiring human review for decisions that meet defined risk criteria

Correct Answer: 4. Requiring human review for decisions that meet defined risk criteria

Explanation:

Human oversight means people remain involved in reviewing or controlling AI behavior when the potential consequences justify additional supervision. For example, an application could automatically process routine requests but require a qualified person to review high-impact decisions or unusual cases. The appropriate level of oversight depends on the intended use, potential risks, and organizational requirements. Human review can help identify errors, unexpected behavior, or situations outside the model’s capabilities. It is an important responsible AI practice because automated systems can make mistakes, and human judgment may be necessary when decisions have significant consequences.

Question 267. What is the primary purpose of content filtering in a generative AI application?

1. To increase the model’s context window
2. To identify and reduce potentially harmful or inappropriate content
3. To create database indexes
4. To convert speech into text

Correct Answer: 2. To identify and reduce potentially harmful or inappropriate content

Explanation:

Content filtering helps an AI application detect and manage content that may violate safety requirements or organizational policies. Depending on the service and configuration, filtering can address categories such as violence, hate, sexual content, or other harmful material. Filters can be applied to user inputs, model outputs, or both. They are especially useful for applications exposed to broad user populations because users may intentionally or unintentionally provide unsafe requests. Content filtering should be considered one part of a broader safety strategy that can also include access controls, monitoring, testing, human oversight, and clearly defined application policies.

Question 268. Which Azure AI Search capability combines different search approaches to improve retrieval?

1. Hybrid search
2. Batch inference
3. Speech synthesis
4. Image segmentation

Correct Answer: 1. Hybrid search

Explanation:

Hybrid search combines traditional keyword-based search with vector-based search to improve the retrieval of relevant information. Keyword search can be effective when exact terms, names, identifiers, or specific phrases are important, while vector search can identify content that is semantically similar even when different words are used. Combining the two approaches can provide stronger results for applications such as enterprise search and retrieval-augmented generation. Azure AI Search supports hybrid retrieval scenarios, allowing developers to use multiple signals when finding relevant documents. This can improve the quality and relevance of information supplied to downstream AI applications.

Question 269. What is the purpose of a vector index in a search solution?

1. To store passwords
2. To convert images into videos
3. To organize numerical representations for similarity search
4. To execute operating-system updates

Correct Answer: 3. To organize numerical representations for similarity search

Explanation:

A vector index is designed to efficiently search numerical representations called embeddings. Embeddings represent the semantic characteristics of content as vectors, allowing a search system to identify items that are mathematically close in vector space. This makes vector search useful for finding documents, products, questions, or other content that is semantically similar rather than relying only on exact keyword matches. In a retrieval-augmented generation solution, a vector index can help retrieve relevant chunks of information that are then supplied to a language model. Efficient vector indexing is therefore an important component of modern semantic retrieval systems.

Question 270. Which Azure AI Language capability can classify text into categories defined for a specific business scenario?

1. Custom text classification
2. Object detection
3. Speech synthesis
4. Image segmentation

Correct Answer: 1. Custom text classification

Explanation:

Custom text classification is an Azure AI Language capability that allows organizations to classify text according to categories relevant to their own business requirements. For example, a company could classify customer messages into categories such as billing, technical support, account access, or product feedback. Unlike a general-purpose classification approach, custom classification can be trained using organization-specific examples and labels. This makes it useful when predefined language features do not match the organization’s terminology or workflow. Developers still need representative training examples and appropriate evaluation to ensure the resulting classifier performs reliably across the intended types of input.

Question 271. Which Azure AI Vision capability can identify general visual characteristics of an image?

1. Image analysis
2. Database replication
3. Text-to-speech
4. SQL indexing

Correct Answer: 1. Image analysis

Explanation:

Azure AI Vision image analysis can examine an image and return information about visual content. Depending on the capabilities being used, image analysis can provide descriptions, tags, detected objects, text, and other visual information. This allows applications to extract useful information from photographs, documents, or other images without building a complete computer vision model from scratch. Developers can use these results in scenarios such as content organization, accessibility, document processing, or image-based search. As with other AI services, applications should evaluate accuracy for their specific image types and consider the consequences of incorrect results.

Question 272. When would a custom Document Intelligence model be useful?

1. When documents follow a unique organizational format that requires specialized extraction
2. When converting speech to audio
3. When searching DNS records
4. When creating network routes

Correct Answer: 1. When documents follow a unique organizational format that requires specialized extraction

Explanation:

A custom Document Intelligence model can be useful when an organization processes documents with layouts, fields, or structures that are specific to its business. Examples may include specialized forms, internal documents, or document types that are not adequately handled by a prebuilt model. The organization can provide representative examples so the model can learn how to identify relevant fields and information. Custom document models are valuable when consistent automated extraction is needed from recurring document formats. Proper training data and evaluation remain important because variations in document layout, quality, and content can affect extraction accuracy.

Question 273. Which factor should be considered when selecting a speech recognition solution?

1. The number of database indexes
2. The language and dialect spoken by users
3. The size of a virtual network
4. The color of the application interface

Correct Answer: 2. The language and dialect spoken by users

Explanation:

Language and dialect support are important considerations when selecting a speech recognition solution because recognition quality can vary depending on the spoken language, accent, dialect, vocabulary, and recording conditions. An application designed for a specific user population should be evaluated using representative speech samples rather than assuming that performance will be identical across all speakers. Background noise, microphone quality, speaking speed, and domain-specific terminology can also affect recognition results. Azure AI Speech provides speech-related capabilities for supported languages and scenarios, but developers should test the service with realistic data before deploying it to production.

Question 274. What should be considered when selecting a text-to-speech voice for an application?

1. The application’s audience, language, and desired voice characteristics
2. The number of database tables
3. The number of network switches
4. The size of a storage account

Correct Answer: 1. The application’s audience, language, and desired voice characteristics

Explanation:

Text-to-speech voice selection should match the application’s users and intended experience. Important factors can include language, locale, pronunciation, speaking style, voice characteristics, and whether the application needs a particular tone or delivery style. For example, an educational application may require a clear and natural voice, while a customer-service application may have different requirements. Developers should also test pronunciation of specialized names, technical terms, and domain-specific vocabulary. Azure AI Speech provides multiple voice options, but selecting an appropriate voice requires evaluating how naturally and clearly it communicates the application’s content to its intended audience.

Question 275. Which Azure Machine Learning capability can expose a trained model for real-time application requests?

1. Online endpoint
2. Data labeling only
3. Notebook comments
4. Model documentation

Correct Answer: 1. Online endpoint

Explanation:

An Azure Machine Learning online endpoint provides a way to expose a deployed machine learning model for real-time inference requests. Applications can send input data to the endpoint and receive predictions with low latency, making this deployment approach suitable for interactive scenarios. Examples include real-time fraud scoring, recommendation requests, or applications that need immediate predictions. Developers need to consider authentication, scaling, monitoring, model versioning, and performance when deploying an endpoint. Online endpoints differ from batch endpoints, which are designed for processing larger amounts of data asynchronously rather than returning an immediate prediction for each individual request.

Question 276. Why is model versioning useful in an AI application?

1. It prevents all model changes
2. It allows teams to track and manage different versions of models
3. It removes the need for testing
4. It automatically guarantees accuracy

Correct Answer: 2. It allows teams to track and manage different versions of models

Explanation:

Model versioning allows teams to identify, track, and manage different iterations of a machine learning or AI model. A new version may result from updated training data, changes to algorithms, improved features, or corrected configuration. Keeping versions organized helps teams understand which model is deployed and compare newer versions against previous ones. It can also support controlled deployment and rollback when a new version does not perform as expected. Versioning does not automatically guarantee that a model is accurate or reliable, so each version should still be evaluated against appropriate datasets and requirements before and after deployment.

Question 277. What is the main purpose of monitoring data drift in a machine learning system?

1. To detect changes in input data characteristics over time
2. To increase storage capacity
3. To encrypt every database column
4. To convert text into speech

Correct Answer: 1. To detect changes in input data characteristics over time

Explanation:

Data drift occurs when the statistical characteristics or distribution of input data change compared with the data used during model development. For example, customer behavior, product preferences, or operating conditions may change over time, causing production inputs to differ from training data. Monitoring for data drift can help teams identify situations where model performance may eventually be affected. Drift does not automatically mean that a model has become unusable, but it is a useful signal for investigation. Teams can examine performance metrics, data quality, and business conditions before deciding whether retraining or other action is necessary.

Question 278. Which machine learning technique is appropriate for grouping similar customers when no predefined customer categories exist?

1. Regression
2. Classification
3. Clustering
4. Speech recognition

Correct Answer: 3. Clustering

Explanation:

Clustering is an unsupervised machine learning technique used to group data points according to similarities without requiring predefined labels. For example, a company may have customer information such as purchasing behavior, spending patterns, or product preferences but may not already know the appropriate customer groups. A clustering algorithm can identify groups with similar characteristics, which analysts can then investigate and interpret. The resulting clusters depend on the selected features and algorithm, so they should not automatically be treated as meaningful business categories without analysis. Clustering is particularly useful for exploratory analysis and discovering patterns in unlabeled data.

Question 279. What problem can class imbalance create in a classification model?

1. The model may perform well overall while performing poorly on a minority class
2. The model cannot contain any features
3. The model automatically becomes a regression model
4. The training dataset is deleted

Correct Answer: 1. The model may perform well overall while performing poorly on a minority class

Explanation:

Class imbalance occurs when some classes in a classification dataset have substantially more examples than others. A model trained on such data may achieve high overall accuracy by predicting the majority class frequently while failing to identify the less common class correctly. This can be especially problematic when the minority class represents an important event, such as fraud or a security incident. Developers should therefore examine metrics such as precision, recall, and the confusion matrix for individual classes rather than relying only on overall accuracy. Appropriate sampling, weighting, additional data, and careful evaluation can help address imbalance.

Question 280. What does a confusion matrix help you evaluate in a classification model?

1. The size of the model’s context window
2. The number of tokens generated
3. The types of correct and incorrect class predictions
4. The amount of cloud storage available

Correct Answer: 3. The types of correct and incorrect class predictions

Explanation:

A confusion matrix summarizes the results of a classification model by comparing predicted classes with actual classes. It can show counts such as true positives, true negatives, false positives, and false negatives for a binary classification problem, with corresponding categories for multiclass problems. These values provide more detail than a single accuracy number and can help calculate metrics such as precision, recall, and F1 score. A confusion matrix is particularly useful when different types of errors have different consequences. Reviewing it can help developers understand where a classifier is performing well and where additional data, tuning, or investigation may be required.