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

 

Question 341

Which capability allows an AI agent to perform an external action?

1. Tool calling
2. OCR
3. Sentiment analysis
4. Embedding

Correct Answer: 4

Explanation

Tool calling allows an AI agent to interact with external functions, APIs, or services to perform actions beyond generating text. For example, an agent may call an order-management API to retrieve information or submit a request. The application should define clear tool descriptions and input schemas so the model understands how each tool should be used. Developers should also validate tool arguments, authenticate requests, enforce authorization, and restrict permissions. Tool calling therefore connects an AI agent with controlled external capabilities while maintaining application-level security.

Question 342

What is the primary purpose of a vector database or vector index?

1. Store passwords
2. Enable similarity-based retrieval
3. Convert speech to text
4. Analyze sentiment

Correct Answer: 2

Explanation

A vector index stores numerical representations of content so applications can perform similarity-based searches. Embeddings represent documents and queries as vectors, allowing the system to identify content that is semantically related even when exact words differ. Vector retrieval is commonly used in RAG applications and semantic knowledge systems. Developers should select an appropriate embedding model and configure the index according to the application’s data and search requirements. Access controls and metadata filtering may also be needed to ensure retrieved information is appropriate for each user.

Question 343

Which service is appropriate for converting written text into natural-sounding speech?

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

Correct Answer: 3

Explanation

Azure AI Speech provides text-to-speech capabilities that can convert written text into synthesized spoken audio. This is useful for voice assistants, accessibility applications, announcements, educational systems, and conversational interfaces. Developers can select supported voices and configure speech characteristics according to the application. Testing should consider pronunciation, language, speaking style, and the clarity of generated audio. Text-to-speech is different from speech-to-text, which converts spoken audio into written text. Both capabilities can be combined in applications that support two-way voice interaction.

Question 344

Why should retrieved content be checked for relevance before generation?

1. Irrelevant context can reduce response quality
2. It increases password security
3. It converts documents into audio
4. It removes the need for indexing

Correct Answer: 1

Explanation

Relevant retrieval is important because generative models use the supplied context when producing an answer. If the retrieved documents are unrelated, outdated, or incomplete, they can confuse the model and contribute to inaccurate responses. Retrieval quality can be improved through appropriate chunking, metadata filters, semantic or vector search, hybrid search, and reranking. Developers should evaluate retrieval separately from generation so they can identify whether a poor answer resulted from incorrect retrieval or from the model’s response generation. High-quality context is an important part of a reliable RAG system.

Question 345

What can metadata filters do in an AI Search solution?

1. Restrict results based on document attributes
2. Convert audio to text
3. Generate passwords
4. Increase model temperature

Correct Answer: 1

Explanation

Metadata filters can restrict search results according to attributes stored with indexed documents. For example, an application may filter documents by department, region, document type, date, or access-related metadata. Filtering can improve relevance and can also support security requirements when combined with proper authorization design. Developers should ensure that metadata is accurate and consistently maintained during indexing. Filters should not be treated as a substitute for complete authorization controls. Applications should still verify that the requesting identity is allowed to access the returned information.

Question 346

Which capability is commonly used to translate spoken language?

1. Object detection
2. Speech translation
3. OCR
4. Vector search

Correct Answer: 2

Explanation

Speech translation allows applications to process spoken language and translate it into another supported language. It can be useful for multilingual conversations, meetings, customer-service applications, and accessibility scenarios. The quality of the result can depend on audio clarity, language support, accents, background noise, and conversational context. Developers should test the capability using realistic audio rather than relying only on ideal recordings. Speech translation differs from standard speech-to-text because its workflow includes translating the recognized spoken content into another language.

Question 347

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

1. Lower latency or cost
2. Automatic administrator access
3. Better document scanning in every case
4. Elimination of monitoring

Correct Answer: 1

Explanation

Smaller AI models can often provide lower latency and lower operating costs for tasks that do not require the capabilities of a larger model. Examples may include simple classification, routing, extraction, or straightforward text-processing tasks. Developers should compare model quality against application requirements instead of selecting a model based only on size. A smaller model may be sufficient for one workload but inadequate for another. Evaluation datasets can help determine whether the model meets accuracy, consistency, safety, and performance requirements before production deployment.

Question 348

Which practice helps protect an agent’s connected tools?

1. Allow every tool without restrictions
2. Validate tool inputs and enforce permissions
3. Disable authentication
4. Trust every model-generated argument

Correct Answer: 2

Explanation

Tool security requires controls around both what an agent can access and what arguments it can send. Applications should validate model-generated tool inputs against expected schemas and business rules before execution. Authentication and authorization should also ensure that the calling identity has permission to perform the requested operation. Tool access should follow least privilege, and sensitive actions may require human approval. Developers should treat model-generated arguments as untrusted input because models can make mistakes or be influenced by malicious instructions in user or retrieved content.

Question 349

Which capability can summarize long text into a shorter version?

1. Text summarization
2. Object detection
3. Speech synthesis
4. Vector indexing

Correct Answer: 3

Explanation

Text summarization produces a shorter representation of longer content while attempting to preserve important information. It can be useful for summarizing reports, customer conversations, support cases, articles, and meeting transcripts. Developers should define the desired summary length, style, and information requirements clearly. Summarization quality should be evaluated because a model may omit important details or introduce unsupported information. For sensitive or high-impact content, applications may need additional validation or human review. Summarization can also be combined with speech, search, and document-processing workflows.

Question 350

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

1. Define high-level behavior and instructions
2. Store vector embeddings
3. Detect objects
4. Encrypt database records

Correct Answer: 4

Explanation

A system message provides high-level instructions that establish how a generative model should behave within an application. It can define the assistant’s role, response requirements, boundaries, formatting expectations, or other behavioral guidance. Developers should keep system instructions clear and avoid exposing unnecessary sensitive information. System instructions are not a replacement for security controls because users or retrieved content may attempt to manipulate model behavior. Authorization, tool restrictions, input validation, and other application-level controls should remain outside the model’s instruction hierarchy.

Question 351

Which search approach can find documents based on conceptual similarity?

1. Vector search
2. OCR
3. Speech recognition
4. Sentiment analysis

Correct Answer: 1

Explanation

Vector search retrieves content based on the similarity between numerical vector representations. Because embeddings capture semantic characteristics, vector search can find documents that express similar ideas even when they do not share the same exact keywords. This makes it useful for natural-language questions and RAG systems. Developers should select an appropriate embedding model and evaluate retrieval results using representative queries. Vector search can also be combined with keyword or semantic search to improve retrieval coverage. Proper filtering and authorization remain necessary when the indexed data contains restricted information.

Question 352

What does authentication establish?

1. What an identity is allowed to do
2. The identity of a requester
3. The relevance of a document
4. The temperature of a model

Correct Answer: 2

Explanation

Authentication establishes or verifies the identity of a user, application, or service making a request. Once identity is established, authorization determines which resources or actions that identity is permitted to access. This distinction is important in AI applications that retrieve private documents or call protected business APIs. Developers should use appropriate authentication mechanisms and avoid relying on the model to identify or authorize users. Authentication credentials should also be protected using secure identity-management practices, and sensitive credentials should not be exposed in prompts or logs.

Question 353

Why is chunk overlap sometimes used when splitting documents?

1. To preserve context between neighboring chunks
2. To increase audio volume
3. To create user accounts
4. To remove all metadata

Correct Answer: 4

Explanation

Chunk overlap repeats a portion of content between neighboring chunks so information near a chunk boundary is less likely to lose important context. Without overlap, a sentence or concept that spans two chunks may be divided in a way that makes retrieval less useful. The appropriate overlap depends on document structure, chunk size, and the retrieval workload. Too much overlap can increase index size and produce redundant context. Developers should test chunking strategies with representative documents and queries to determine whether overlap improves retrieval quality.

Question 354

Which capability can identify the main concepts in a piece of text?

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

Correct Answer: 2

Explanation

Key phrase extraction identifies significant words or phrases that represent important concepts in text. It can support document categorization, content discovery, search, analytics, and customer-feedback processing. Unlike sentiment analysis, which focuses on the expressed opinion or emotional tone, key phrase extraction focuses on important concepts and topics. Developers should test extracted phrases against representative content because terminology varies between domains. For automated business workflows, extracted information should be validated when incorrect phrases could cause incorrect categorization or other downstream actions.

Question 355

What should an application do before allowing an agent to access private data?

1. Verify authorization
2. Increase temperature
3. Remove all filters
4. Disable identity checks

Correct Answer: 2

Explanation

Authorization should be verified before an agent retrieves or operates on private data. The application should determine whether the authenticated identity has permission to access the requested resource or perform the requested action. This control should be enforced by the application or connected service rather than relying on the model’s instructions. Access filters, role-based permissions, metadata, and resource-level controls can support this process. Developers should test access using identities with different permissions to ensure the agent cannot retrieve information that the requesting user is not authorized to see.

Question 356

Which capability helps evaluate whether a generated answer is supported by retrieved information?

1. Groundedness evaluation
2. OCR
3. Object detection
4. Speech synthesis

Correct Answer: 1

Explanation

Groundedness evaluation checks whether generated responses are supported by the information provided to the model, such as retrieved documents or other trusted context. This is especially important in RAG applications because a response can sound convincing while containing information that is not supported by the retrieved sources. Developers can use evaluation datasets and defined criteria to measure groundedness consistently. Groundedness evaluation should be combined with retrieval-quality testing and factual validation where appropriate. It helps identify unsupported generation but does not guarantee complete correctness.

Question 357

Which capability can extract text from a scanned document image?

1. OCR
2. Semantic ranking
3. Text-to-speech
4. Rate limiting

Correct Answer: 1

Explanation

OCR, or Optical Character Recognition, extracts machine-readable text from images and scanned documents. It is useful when a document does not contain an accessible digital text layer. OCR can serve as an initial step in document search, information extraction, and automated processing workflows. Results can vary depending on image quality, document orientation, font style, handwriting, and background conditions. Developers should validate extracted text when it is used to populate important fields or trigger business operations. OCR can also be combined with Document Intelligence for richer document analysis.

Question 358

Why should an AI application monitor token usage?

1. To understand usage and manage costs
2. To improve camera resolution
3. To identify physical objects
4. To replace authorization

Correct Answer: 2

Explanation

Monitoring token usage helps developers understand how much model input and output their application consumes. Token usage can affect operating costs, throughput, and sometimes latency. Monitoring can reveal unusually large prompts, excessive retrieved context, inefficient conversations, or unexpected usage patterns. Developers can use this information to optimize prompts, chunk sizes, retrieval results, and model selection. Token monitoring does not replace quality or security monitoring. A production AI system should consider usage alongside response quality, errors, latency, safety, and other operational metrics.

Question 359

What is the purpose of content filtering in an AI application?

1. Reduce harmful or inappropriate content
2. Increase vector dimensions
3. Translate every document
4. Create database indexes

Correct Answer: 3

Explanation

Content filtering can help detect or restrict certain harmful, unsafe, or inappropriate inputs and outputs in an AI application. Filters can be part of a broader responsible AI strategy that also includes system instructions, input validation, access controls, monitoring, and human review where appropriate. Developers should understand the categories and thresholds supported by the selected service and test realistic application scenarios. Content filtering should not be treated as the only security mechanism because harmful behavior can involve tools, authorization, data access, and prompt injection as well.

Question 360

Which practice helps maintain AI application quality after deployment?

1. Continuous evaluation and monitoring
2. Removing all test cases
3. Disabling telemetry
4. Avoiding model comparisons

Correct Answer: 4

Explanation

Continuous evaluation and monitoring help developers detect changes in AI application quality after deployment. Real-world data and user behavior can differ from development and test scenarios, and model or data changes can introduce unexpected regressions. Monitoring can track quality indicators, latency, errors, token usage, safety-related signals, and retrieval performance. Evaluation datasets can provide consistent comparisons between versions. Developers should establish thresholds and review results regularly so problems can be identified and addressed rather than relying only on the initial pre-deployment evaluation.