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

Question 341. Which Azure AI Foundry capability can help developers evaluate a generative AI application using defined criteria?

1. Evaluation
2. Azure DNS
3. Azure Storage Explorer
4. Network routing

Correct Answer: 1. Evaluation

Explanation:

Evaluation helps developers assess whether a generative AI application meets defined quality, safety, and performance requirements. Depending on the scenario, evaluation can examine factors such as relevance, groundedness, coherence, factual accuracy, safety, or task-specific success. Developers can use representative test inputs to compare different prompts, models, or application configurations. Evaluation is especially important because a model that produces fluent responses is not necessarily reliable or suitable for a particular business purpose. Establishing measurable criteria before testing makes it easier to identify weaknesses and determine whether an AI application is ready for broader use.

Question 342. What is groundedness in a generative AI response?

1. The response is supported by the provided source information
2. The response contains the maximum number of tokens
3. The model uses the largest available server
4. The response contains no punctuation

Correct Answer: 1. The response is supported by the provided source information

Explanation:

Groundedness describes how well a generated response is supported by the information provided to the model as context. In a retrieval-augmented generation application, for example, the model may receive information retrieved from an organization’s documents. A grounded response should rely on that supplied information rather than introducing unsupported claims. Evaluating groundedness can help identify situations where a model produces statements that are not supported by its retrieved sources. It is important to remember that grounding does not automatically guarantee that the source information itself is correct, current, or appropriate. Source quality and retrieval quality must also be evaluated.

Question 343. Which technique can reduce the amount of irrelevant information provided to a language model?

1. Retrieving only relevant documents or document chunks
2. Sending every available document with every request
3. Removing all search functionality
4. Disabling document indexing

Correct Answer: 1. Retrieving only relevant documents or document chunks

Explanation:

Retrieving only relevant information helps an AI application provide a language model with focused context rather than overwhelming it with unnecessary content. This is especially important in retrieval-augmented generation systems where an application may have access to thousands or millions of documents. A search system can identify relevant documents or chunks based on the user’s query, and only selected content can then be included in the model request. Better retrieval can improve relevance, reduce unnecessary token usage, and help the model focus on useful evidence. Developers should evaluate retrieval quality because missing relevant information can also reduce response quality.

Question 344. What is chunking in a retrieval-augmented generation solution?

1. Dividing large documents into smaller pieces for retrieval
2. Encrypting an entire database
3. Converting text into speech
4. Removing all document metadata

Correct Answer: 1. Dividing large documents into smaller pieces for retrieval

Explanation:

Chunking divides large documents into smaller sections so that individual pieces can be indexed and retrieved more efficiently. In a retrieval-augmented generation system, a user question may relate to only one section of a large document. Providing the entire document to the model may be inefficient or exceed context limits. By creating smaller chunks, the search system can retrieve the sections most relevant to a query. Chunk size and overlap should be selected carefully because chunks that are too small may lose important context, while chunks that are too large may reduce retrieval precision and increase unnecessary information in the model prompt.

Question 345. Why can metadata be useful in an enterprise search solution?

1. It can provide additional information for filtering and organizing search results
2. It automatically trains every AI model
3. It replaces authentication
4. It converts documents into audio

Correct Answer: 1. It can provide additional information for filtering and organizing search results

Explanation:

Metadata provides descriptive information about content, such as document title, author, department, date, category, or access classification. In an enterprise search solution, metadata can help users and applications filter and organize results more effectively. For example, a search application might restrict results to documents from a particular department or documents updated within a certain period. Metadata can also support security-aware retrieval when appropriate permissions are associated with records. Developers should ensure that metadata is accurate and maintained consistently because incorrect metadata can reduce search quality or potentially expose inappropriate information if it is used incorrectly.

Question 346. Which approach can improve the security of a RAG application that uses private company documents?

1. Enforcing authorization before retrieving documents
2. Making every document publicly accessible
3. Removing all user authentication
4. Sending all company documents to every user

Correct Answer: 1. Enforcing authorization before retrieving documents

Explanation:

Authorization should be enforced before private documents are retrieved and supplied to a generative AI model. A RAG application should not assume that because a document is searchable, every user is allowed to access it. The application should determine the user’s identity and permissions and then retrieve only information that the user is authorized to see. This reduces the risk of confidential information appearing in model responses. Additional controls such as secure storage, encryption, auditing, filtering, and careful prompt construction can further strengthen the solution. Security should be implemented at the application and data layers rather than relying on the model alone.

Question 347. Which Azure AI Content Safety capability can help identify potentially harmful text?

1. Text content analysis
2. Model versioning
3. Data labeling
4. Compute scaling

Correct Answer: 1. Text content analysis

Explanation:

Azure AI Content Safety provides capabilities for analyzing potentially harmful content in text and other supported inputs. Content analysis can help applications identify categories of content that may violate defined safety requirements or organizational policies. This can be useful for user-generated content, chat applications, assistants, and other systems that interact with unrestricted input. Developers can use safety analysis as one layer in a broader protection strategy that may include authentication, authorization, monitoring, human review, and application-specific rules. Content safety services do not replace all application security controls, and teams should test them against scenarios relevant to their intended users and risks.

Question 348. Which responsible AI principle emphasizes assigning responsibility for the behavior and outcomes of an AI system?

1. Accountability
2. Compression
3. Storage optimization
4. Network segmentation

Correct Answer: 1. Accountability

Explanation:

Accountability means that appropriate people or organizations remain responsible for an AI system and its outcomes. This includes establishing ownership, defining responsibilities, documenting important decisions, monitoring system behavior, and providing processes for addressing problems. AI systems should not be treated as independent decision makers without human or organizational responsibility. Accountability is especially important when an AI system affects users, customers, employees, or other stakeholders. Organizations can support accountability through governance processes, documentation, monitoring, testing, escalation procedures, and clear ownership. The exact controls depend on the application’s purpose, risks, and organizational requirements.

Question 349. What is a benefit of using representative evaluation data?

1. It provides a more realistic indication of performance in the intended scenario
2. It guarantees perfect accuracy
3. It eliminates the need for monitoring
4. It prevents all model drift

Correct Answer: 1. It provides a more realistic indication of performance in the intended scenario

Explanation:

Representative evaluation data resembles the types of inputs the AI system is expected to encounter after deployment. Using such data provides a more realistic indication of how the model may perform for its intended users and scenarios. If evaluation data is too simple, too small, or substantially different from production inputs, performance measurements may not accurately reflect real-world behavior. Evaluation should also include relevant edge cases and important user groups where appropriate. Representative data does not guarantee perfect performance, but it helps teams make better-informed decisions about model quality, limitations, and readiness for deployment.

Question 350. Which metric is commonly used to evaluate a binary classification model by balancing precision and recall?

1. F1 score
2. Mean Squared Error
3. Mean Absolute Error
4. Storage utilization

Correct Answer: 1. F1 score

Explanation:

The F1 score combines precision and recall into a single metric by calculating their harmonic mean. It can be useful when both false positives and false negatives matter and a balance between precision and recall is desired. A model with high precision but poor recall, or high recall but poor precision, may have a lower F1 score than a model that balances both measures effectively. F1 should not automatically be considered the best metric for every classification problem. Developers should select metrics based on the business consequences of different errors and may need to examine precision, recall, accuracy, and class-specific results as well.

Question 351. Which Azure Machine Learning capability can automatically explore multiple machine learning algorithms and configurations?

1. Automated ML
2. Azure DNS
3. Azure Blob Storage only
4. Azure Firewall

Correct Answer: 1. Automated ML

Explanation:

Automated Machine Learning, or Automated ML, can automate parts of the machine learning development process by exploring different algorithms, preprocessing approaches, and model configurations for supported tasks. This can help developers identify promising models without manually testing every possible combination. Automated ML can be useful for scenarios such as classification, regression, and other supported machine learning tasks. However, automation does not remove the need for understanding the business problem, preparing appropriate data, selecting relevant evaluation metrics, and validating the resulting model. Developers should review the generated models and ensure that the selected solution is appropriate for the intended application.

Question 352. What is the primary purpose of model interpretability?

1. To help stakeholders understand how a model produces or supports its predictions
2. To increase storage capacity
3. To remove all model parameters
4. To convert speech into text

Correct Answer: 1. To help stakeholders understand how a model produces or supports its predictions

Explanation:

Model interpretability helps developers and stakeholders understand factors that contribute to model predictions or decisions. This can be important when users need to assess whether a model is behaving appropriately or when an organization needs to investigate unexpected results. The level of interpretability required depends on the model, application, and risks involved. Some models are easier to interpret than others, and explanation techniques may provide different types of information rather than a complete description of every internal computation. Interpretability should therefore be considered alongside evaluation, transparency, documentation, and human oversight when designing responsible AI systems.

Question 353. Which practice can help identify whether a model performs differently across user groups?

1. Evaluating performance separately for relevant groups
2. Measuring only overall accuracy
3. Removing group information from all evaluation
4. Testing with one example

Correct Answer: 1. Evaluating performance separately for relevant groups

Explanation:

Evaluating performance separately across relevant groups can reveal differences that may be hidden by an overall performance metric. For example, a model could have strong overall accuracy while producing higher error rates for a particular population. Group-based evaluation can examine metrics such as precision, recall, accuracy, or other task-specific measurements across relevant groups. This information can help developers identify potential fairness concerns and investigate whether differences are caused by data representation, model behavior, or other factors. The groups and metrics used should be appropriate to the application’s context, and evaluation should respect applicable privacy and governance requirements.

Question 354. Which action can help improve the quality of training data?

1. Removing or correcting inaccurate and inconsistent records
2. Adding random values to every record
3. Deleting all labels
4. Ignoring missing values without analysis

Correct Answer: 1. Removing or correcting inaccurate and inconsistent records

Explanation:

Data quality has a direct effect on machine learning performance because models learn patterns from the data provided during training. Inaccurate, duplicated, inconsistent, incomplete, or incorrectly labeled records can introduce noise and cause a model to learn inappropriate relationships. Data preparation may include identifying errors, handling missing values, removing inappropriate duplicates, correcting labels, and ensuring consistent formats. The appropriate treatment depends on the specific dataset and business problem. Developers should avoid blindly deleting records because some unusual values may represent legitimate cases. Data quality processes should be documented and validated to ensure that useful information is not unintentionally removed.

Question 355. What is concept drift?

1. A change in the relationship between input features and the target outcome over time
2. A change in the size of a storage account
3. A change in the model’s file name
4. A change in network bandwidth

Correct Answer: 1. A change in the relationship between input features and the target outcome over time

Explanation:

Concept drift occurs when the relationship between input variables and the target outcome changes over time. For example, customer behavior that previously indicated a high likelihood of purchasing a product may become less predictive after market conditions change. This differs from data drift, which generally refers to changes in the distribution of input data. Concept drift can cause a previously effective model to become less accurate because the patterns it learned are no longer valid. Monitoring model performance and relevant data can help organizations identify possible drift and determine whether retraining or other updates are necessary.

Question 356. Which situation is an example of data drift rather than concept drift?

1. The distribution of customer ages in production changes substantially compared with training data
2. The meaning of customer behavior changes over time
3. The relationship between features and the target changes
4. The business definition of success changes

Correct Answer: 1. The distribution of customer ages in production changes substantially compared with training data

Explanation:

Data drift occurs when the distribution or characteristics of input data change between the data used for model development and the data observed in production. For example, a model may have been trained using customers mostly between ages 20 and 50, while the production population later contains a much larger proportion of older customers. The relationship between inputs and outcomes may remain unchanged, but the input distribution has shifted. Monitoring data characteristics can help identify such changes. Data drift does not automatically mean the model has failed, but it can be an important signal that further investigation is needed.

Question 357. Which Azure AI Search feature can use filters to restrict results based on document attributes?

1. Metadata filtering
2. Speech synthesis
3. Model training
4. Image segmentation

Correct Answer: 1. Metadata filtering

Explanation:

Metadata filtering allows search applications to restrict results according to attributes associated with indexed documents. For example, an application could filter documents by department, category, language, publication date, or another supported metadata field. This can improve the usefulness of search results by narrowing the result set to information that matches the user’s requirements. In enterprise scenarios, filtering can also be part of a broader security-aware retrieval design, although authorization should not depend solely on ordinary search filtering. Developers should ensure that metadata values are accurate and that access restrictions are enforced appropriately before information is provided to users or AI models.

Question 358. Why might a developer choose a smaller generative AI model instead of a larger model?

1. It may provide lower latency or cost for a task that does not require the larger model’s capabilities
2. Smaller models always produce more accurate answers
3. Smaller models never require evaluation
4. Smaller models automatically understand every domain

Correct Answer: 1. It may provide lower latency or cost for a task that does not require the larger model’s capabilities

Explanation:

A smaller generative AI model may be appropriate when the application does not require the capabilities of a larger model. Smaller models can potentially provide advantages such as lower latency, lower resource requirements, or lower operating costs, depending on the service and workload. However, model selection should be based on measured performance for the intended task rather than model size alone. A larger model may provide stronger performance for complex tasks, while a smaller model may be sufficient for simpler workloads. Developers should compare candidate models using representative prompts, quality requirements, latency expectations, and cost considerations.

Question 359. What should a developer do if an AI model produces unsafe responses during testing?

1. Investigate the cause and apply appropriate safeguards before deployment
2. Ignore the responses because testing is optional
3. Deploy the model immediately
4. Remove all monitoring

Correct Answer: 1. Investigate the cause and apply appropriate safeguards before deployment

Explanation:

Unsafe responses identified during testing should be treated as evidence that additional investigation or safeguards may be required. Developers should determine the circumstances that caused the behavior and evaluate whether the issue is related to prompts, model capabilities, retrieved content, input handling, configuration, or another part of the application. Possible safeguards can include content filtering, improved instructions, input validation, access controls, output checks, monitoring, or human review depending on the scenario. Testing should then be repeated to determine whether the safeguards reduce the identified risk. An unsafe behavior should not simply be ignored because it occurred during testing rather than production.

Question 360. Which approach best supports responsible deployment of an AI solution?

1. Define requirements, evaluate the system, apply safeguards, and monitor it after deployment
2. Deploy first and evaluate only if users complain
3. Remove all human oversight
4. Use the same model configuration for every scenario without testing

Correct Answer: 1. Define requirements, evaluate the system, apply safeguards, and monitor it after deployment

Explanation:

Responsible AI deployment involves multiple stages rather than treating deployment as the final technical step. Teams should first understand the intended purpose, users, risks, and requirements of the application. The system should then be evaluated using representative data and relevant quality and safety criteria. Appropriate safeguards can include access controls, content filtering, human oversight, monitoring, and data protection measures. After deployment, ongoing monitoring can identify changes in performance, input data, or system behavior. This lifecycle approach helps organizations manage the limitations and risks of AI systems while maintaining evidence about whether the solution continues to meet its intended requirements.