View Full Microsoft AI-103 Exam Dumps and Practice Test Dumps.

 

Question 201

What is the main purpose of chunking documents in a RAG system?

1. To convert speech into text
2. To divide large content into searchable sections
3. To remove authentication
4. To increase image resolution

Correct Answer: 2

Explanation

Chunking divides large documents into smaller sections that can be indexed, retrieved, and supplied to a language model as context. Appropriate chunk sizes can improve retrieval because the system can identify specific passages instead of returning an entire document. Chunking should preserve enough surrounding information for each section to remain meaningful. Developers may also use overlap between chunks when important information could span boundaries. The best configuration depends on document structure, content type, embedding model, and the application’s retrieval requirements.

Question 202

Which Azure service is designed to analyze images and extract visual information?

1. Azure AI Vision
2. Azure AI Speech
3. Azure AI Search
4. Azure AI Language

Correct Answer: 1

Explanation

Azure AI Vision provides capabilities for analyzing visual content. Depending on the supported feature, an application can use it for tasks such as image analysis, optical character recognition, and object-related visual processing. This makes it useful for applications that need to understand photographs, scanned images, or other visual content. Developers should select the appropriate vision capability based on the specific requirement. Testing with representative images is important because image quality, lighting, resolution, and content can affect analysis results.

Question 203

Why is reranking useful in a search-based AI application?

1. It converts text into speech
2. It stores application secrets
3. It improves the ordering of retrieved results
4. It creates user accounts

Correct Answer: 3

Explanation

Reranking evaluates an initial set of retrieved results and reorganizes them according to their relevance to the user’s query. This can improve the quality of context provided to a generative model. A search system may first retrieve a relatively broad group of candidates and then use a reranking method to place the most relevant passages higher in the results. Better ranking can reduce irrelevant context and improve grounded responses. Developers should evaluate reranking with representative queries rather than assuming it always improves every workload.

Question 204

Which authentication approach can avoid storing long-lived credentials in application code?

1. Plain-text passwords
2. Managed identity
3. Hard-coded API keys
4. Shared administrator accounts

Correct Answer: 2

Explanation

Managed identity allows an Azure resource to authenticate to supported services without requiring developers to place long-lived credentials directly in application code. The identity can be granted only the permissions required for the workload. This reduces the need to store and rotate certain credentials manually and can improve security compared with hard-coded secrets. Developers must still configure appropriate authorization and role assignments. Managed identity is particularly useful for Azure-hosted applications that need secure access to services such as storage, databases, or other supported resources.

Question 205

Which feature is useful when an AI application must return consistent fields such as name, date, and amount?

1. Object detection
2. Speech translation
3. Structured output
4. Image segmentation

Correct Answer: 3

Explanation

Structured output is useful when an application requires predictable fields and a defined response format. Instead of receiving unrestricted natural-language text, the model can be instructed or configured to produce data that follows a specified structure. For example, an invoice-processing application might require fields for vendor name, invoice date, and total amount. The application should still validate the returned values and structure before using them. Structured output is especially valuable when model responses are passed directly to downstream application components.

Question 206

What does semantic search focus on?

1. Meaning and intent
2. File size only
3. Audio volume
4. Image dimensions

Correct Answer: 1

Explanation

Semantic search focuses on the meaning and intent of a query rather than relying only on exact keyword matches. It can identify content that is conceptually related even when the wording differs from the user’s search terms. This makes semantic search useful for knowledge bases, enterprise documents, and question-answering applications. Search quality depends on the configuration, content, and supported ranking capabilities. Developers should compare search results against representative queries to determine whether semantic retrieval is producing relevant information for their particular workload.

Question 207

A chatbot must answer questions using private employee documents. What should be enforced before retrieval?

1. Public access for everyone
2. Authorization and access controls
3. Removal of document permissions
4. Anonymous administrator access

Correct Answer: 2

Explanation

Private employee documents should be protected by authorization and access controls before information is retrieved and presented to a user. A RAG system should not assume that because a document exists in a search index, every user is allowed to receive it. Access rules can be applied using identity information, document permissions, metadata filters, or other appropriate controls. The application should verify authorization before exposing sensitive content. Security testing should also confirm that users cannot bypass these restrictions through alternative queries or prompts.

Question 208

What is a key benefit of using a smaller AI model for simple tasks?

1. Higher microphone sensitivity
2. Lower latency and cost
3. Automatic authorization
4. Larger document storage

Correct Answer: 2

Explanation

Smaller AI models can be useful for tasks that do not require the capabilities of a larger model. They often require fewer computational resources, which can reduce latency and operating costs. Examples include simple classification, routing, extraction, or straightforward transformations. However, model size should not be selected based on cost alone. Developers should evaluate accuracy, reasoning requirements, language support, context needs, and production performance. A smaller model is appropriate when it can meet the application’s quality requirements while providing operational benefits.

Question 209

Which capability converts written text into spoken audio?

1. Speech-to-text
2. Text-to-speech
3. Language detection
4. Named entity recognition

Correct Answer: 2

Explanation

Text-to-speech converts written text into spoken audio. It can be used in applications such as virtual assistants, accessibility features, automated announcements, and voice-enabled customer experiences. Developers can provide text to a supported speech service and receive generated audio in an appropriate format. Speech quality can depend on the selected voice, language, pronunciation requirements, and configuration. Applications should also consider latency and user experience when generating audio dynamically. Text-to-speech is different from speech-to-text, which performs the reverse transformation.

Question 210

Why should AI-generated tool arguments be validated?

1. Models can produce incorrect or unsafe values
2. Validation increases screen brightness
3. Tool arguments are always guaranteed to be correct
4. Validation replaces authentication

Correct Answer: 1

Explanation

AI models can generate incorrect, incomplete, or unexpected tool arguments, so applications should validate those arguments before executing an operation. For example, an order-management tool should verify that an order identifier exists and that requested quantities are valid. Validation can prevent malformed requests from reaching downstream systems. It does not replace authentication or authorization, because a technically valid request may still be unauthorized. Developers should apply schema validation, business-rule checks, and permission checks before allowing an agent to perform sensitive operations.

Question 211

What is the purpose of an Azure AI Search indexer?

1. Automatically ingest and index content from supported data sources
2. Convert speech to audio
3. Create user passwords
4. Generate video recordings

Correct Answer: 1

Explanation

An Azure AI Search indexer can automate the process of retrieving content from supported data sources and adding it to a search index. This reduces the need for an application to manually process every source record and submit it to the index. Depending on the configuration, indexers can work with supported storage and data sources and can participate in document-processing workflows. Developers should configure the data source, index, field mappings, and schedule appropriately while considering permissions, freshness requirements, and processing failures.

Question 212

Which Azure AI service can perform named entity recognition on text?

1. Azure AI Vision
2. Azure AI Language
3. Azure AI Speech
4. Azure AI Search

Correct Answer: 2

Explanation

Azure AI Language provides natural language processing capabilities including named entity recognition. NER identifies entities such as people, organizations, locations, dates, and other supported categories within text. This can help applications extract structured information from emails, customer messages, reports, and documents. The extracted entities can then be stored, searched, classified, or used in downstream workflows. Developers should evaluate recognition quality using representative text because terminology, formatting, language, and domain-specific content can influence the accuracy of extracted entities.

Question 213

Which approach can improve retrieval when users phrase questions differently from indexed documents?

1. Query rewriting
2. Image resizing
3. Audio compression
4. Password rotation

Correct Answer: 1

Explanation

Query rewriting can transform a user’s original question into a form that is more suitable for retrieval. This can be useful when users use informal wording, ambiguous terms, or language that differs from the terminology found in indexed documents. A rewritten query may help search components identify more relevant content. However, developers should ensure that rewriting does not change the user’s intended meaning. Evaluation with real queries is important because an inaccurate rewrite can reduce retrieval quality rather than improve it.

Question 214

What does OCR primarily extract from an image or scanned document?

1. Spoken audio
2. Written text
3. User permissions
4. Vector embeddings only

Correct Answer: 2

Explanation

Optical character recognition, or OCR, extracts machine-readable text from images and scanned documents. It is useful when information exists as a visual representation rather than as digitally selectable text. OCR can support workflows such as document search, form processing, invoice analysis, and information extraction. Accuracy can be affected by image quality, handwriting, fonts, orientation, and complex layouts. Developers should test OCR using representative documents and may combine OCR with document-analysis capabilities when they need both text and structural information.

Question 215

Which strategy helps an AI application continue operating when a preferred model becomes temporarily unavailable?

1. Model fallback
2. Removing monitoring
3. Disabling authentication
4. Increasing prompt length

Correct Answer: 1

Explanation

Model fallback allows an application to use an alternative model or deployment when the preferred option cannot process a request. This can improve resilience when a model is temporarily unavailable, capacity is constrained, or a service condition prevents normal processing. The fallback model should still satisfy the application’s minimum quality and security requirements. Developers should define clear fallback conditions and monitor how often fallback occurs. Testing is important because different models may produce different outputs, latency, costs, or capabilities.

Question 216

What is a major purpose of content filtering in generative AI applications?

1. Help identify or block inappropriate content
2. Increase database storage
3. Convert images into vectors automatically
4. Replace all authorization checks

Correct Answer: 1

Explanation

Content filtering can help identify or block categories of potentially harmful or inappropriate input and output in generative AI applications. It provides an additional safety layer around model interactions and can be configured according to supported categories and application requirements. Content filtering should not be treated as the only security control. Applications should also use authentication, authorization, input validation, monitoring, and appropriate business rules. Developers should test filtering behavior with representative scenarios and understand that automated safety systems can have limitations.

Question 217

Which search approach combines keyword matching with vector-based retrieval?

1. OCR search
2. Hybrid search
3. Audio search
4. Image segmentation

Correct Answer: 2

Explanation

Hybrid search combines traditional keyword-based retrieval with vector-based semantic retrieval. Keyword matching can be useful when exact terms, product names, identifiers, or specific phrases matter, while vector search can identify content based on semantic similarity. Combining both approaches can provide broader retrieval coverage for many enterprise scenarios. Developers should configure ranking and weighting appropriately and evaluate the results using representative queries. Hybrid search is particularly useful when a knowledge base contains both natural-language content and precise terms that users may search for directly.

Question 218

Why should application logs avoid unnecessary sensitive information?

1. To reduce accidental exposure of private data
2. To make models larger
3. To increase microphone volume
4. To remove the need for encryption

Correct Answer: 1

Explanation

Application logs can contain information that becomes accessible to operators, monitoring systems, or other services, so unnecessary sensitive data should not be recorded. Data minimization helps reduce the impact of accidental exposure and supports better privacy practices. Developers should determine which information is genuinely required for troubleshooting and monitoring. Sensitive values may need to be removed, masked, or protected. Logging policies should also consider retention periods and access permissions so that diagnostic information does not become an unnecessary source of data exposure.

Question 219

What is the purpose of human-in-the-loop approval for an AI agent?

1. Allow a person to review selected actions before execution
2. Increase the model’s token limit automatically
3. Replace all search indexes
4. Convert speech into text

Correct Answer: 1

Explanation

Human-in-the-loop approval allows a person to review and approve selected AI-generated actions before they are executed. This is especially useful for operations that could have significant financial, legal, operational, or data-related consequences. For example, an agent might prepare a transaction or account change but require human confirmation before completing it. Developers should define which actions require approval and ensure the approval process provides enough information for meaningful review. Human approval complements, rather than replaces, authentication, authorization, and validation controls.

Question 220

Which practice helps determine whether a RAG application is producing grounded answers?

1. Evaluate responses against representative questions and source information
2. Disable document retrieval
3. Increase temperature without testing
4. Remove source documents after indexing

Correct Answer: 1

Explanation

Grounding evaluation checks whether generated responses are supported by the information retrieved from approved knowledge sources. Developers can create representative questions with expected source information and evaluate whether the system retrieves relevant evidence and uses it accurately. Metrics or review criteria can assess factors such as relevance, correctness, citation quality, and unsupported claims. Evaluation should include common cases and difficult edge cases. Regular testing is important because changes to documents, prompts, models, retrieval settings, or indexing can affect grounding quality over time.