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

 

Question 321

Which component stores documents so they can be searched by an AI application?

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

Correct Answer: 1

Explanation

An Azure AI Search index stores searchable representations of data so applications can retrieve relevant information efficiently. In a RAG solution, documents are processed, indexed, and then searched when a user submits a question. The retrieved content can be supplied to a generative model as context. Indexes can support capabilities such as keyword, vector, semantic, and hybrid search. Developers should design fields, searchable content, metadata, and access-control information according to the application’s retrieval and security requirements.

Question 322

What is the main purpose of embeddings?

1. Convert speech into audio
2. Represent content as numerical vectors
3. Store passwords
4. Detect objects in images

Correct Answer: 2

Explanation

Embeddings represent text or other supported content as numerical vectors that capture semantic characteristics. Similar content can produce vectors that are closer together in vector space, allowing applications to perform similarity-based retrieval. Embeddings are commonly used in RAG systems and semantic search solutions. Developers should select an embedding model appropriate for the language, content, and application requirements. The resulting vectors can be stored in a vector-capable search system and retrieved based on similarity to a user’s query.

Question 323

Which technique can help protect an AI agent from malicious instructions inside retrieved documents?

1. Increasing temperature
2. Adding more documents
3. Prompt injection defenses
4. Removing authentication

Correct Answer: 3

Explanation

Prompt injection defenses help reduce the risk that untrusted content will manipulate an AI model into ignoring its intended instructions or performing unauthorized actions. Retrieved documents, web pages, emails, and tool outputs should be treated as potentially untrusted data rather than trusted instructions. Applications can separate trusted instructions from external content, restrict tool permissions, validate tool arguments, and apply additional safety controls. Security testing should include malicious and adversarial inputs because prompt injection cannot be addressed reliably through a single defensive technique.

Question 324

Why is document chunking used in RAG systems?

1. To divide large content into searchable sections
2. To increase audio volume
3. To authenticate users
4. To detect faces

Correct Answer: 1

Explanation

Document chunking divides large documents into smaller sections that can be indexed and retrieved individually. Smaller chunks can make retrieval more precise because the application can return the portions most relevant to a user’s question instead of sending an entire document to the model. Chunk size and overlap should be selected according to the document structure and expected queries. Poor chunking can separate related information or produce excessive irrelevant context. Developers should evaluate different strategies using representative documents and retrieval queries.

Question 325

Which feature helps search results consider the meaning of a query?

1. Semantic search
2. OCR
3. Speech synthesis
4. Key Vault

Correct Answer: 1

Explanation

Semantic search uses language understanding to identify content that is conceptually related to a query rather than relying only on exact keyword matches. This can help when users phrase questions differently from the wording used in stored documents. Semantic search is particularly useful for natural-language queries in knowledge bases and RAG applications. It can complement keyword and vector retrieval depending on the solution design. Developers should evaluate semantic search using realistic queries because relevance depends on the quality, structure, and content of the indexed data.

Question 326

What is a major advantage of asynchronous processing?

1. It can handle long-running tasks without blocking the requester
2. It removes all security requirements
3. It guarantees perfect model responses
4. It eliminates the need for monitoring

Correct Answer: 1

Explanation

Asynchronous processing allows an application to start a long-running operation and continue other work while the operation completes. This is useful for tasks such as large document processing, batch analysis, or other operations that may take significant time. The application can track the operation and retrieve the result when processing is finished. Developers should design appropriate status handling, retries, timeouts, and failure paths. Asynchronous processing can improve application responsiveness, but it does not automatically solve scalability, reliability, or security requirements.

Question 327

Which Azure service is commonly used to extract structured fields from documents?

1. Azure AI Speech
2. Azure AI Vision
3. Azure AI Document Intelligence
4. Azure AI Search

Correct Answer: 3

Explanation

Azure AI Document Intelligence is designed to analyze documents and extract information from supported document types. It can identify text, layout elements, tables, and fields depending on the model and document scenario. This makes it useful for invoices, receipts, forms, and other business documents. Developers should select an appropriate prebuilt or custom model and validate extracted values before using them in important workflows. Document quality, layout variation, handwriting, and unusual formatting can affect extraction accuracy and should be included in testing.

Question 328

What does least privilege mean for an AI agent?

1. Give the agent only the permissions it needs
2. Give the agent administrator access
3. Disable all authentication
4. Allow every available tool

Correct Answer: 1

Explanation

Least privilege means giving an identity, application, or AI agent only the permissions required to perform its intended tasks. An agent that only needs to read order information should not automatically receive permissions to modify accounts or delete records. Applying least privilege reduces the potential impact of compromised credentials, malicious inputs, or unintended actions. Developers should review tool permissions, resource access, and identities regularly. Authorization should be enforced by the application and connected services rather than relying on the model to decide whether an operation is permitted.

Question 329

Which capability identifies people, organizations, locations, or other named entities in text?

1. Sentiment analysis
2. Named entity recognition
3. Text-to-speech
4. Image classification

Correct Answer: 2

Explanation

Named entity recognition identifies specific types of entities in text, such as people, organizations, locations, dates, or other supported categories. It can help applications organize documents, extract structured information, improve search, and analyze business content. For example, an application processing news articles could identify company and location names automatically. Entity categories and accuracy depend on the service and model being used. Developers should evaluate results against representative data, especially when extracted entities will be used for automated decisions or downstream processing.

Question 330

What does temperature primarily control in generative AI output?

1. Database permissions
2. Network bandwidth
3. Output randomness
4. Document indexing

Correct Answer: 3

Explanation

Temperature is a generation parameter that influences how predictable or varied model outputs can be. Lower values generally encourage more consistent and focused responses, while higher values can produce greater variation. The appropriate setting depends on the application. Tasks requiring predictable outputs may benefit from lower randomness, while creative generation may allow more variation. Temperature does not improve factual accuracy by itself and cannot replace grounding, retrieval, evaluation, or output validation. Developers should test settings with representative prompts before selecting a configuration for production.

Question 331

Why should an AI application use access filters on private documents?

1. To prevent unauthorized retrieval
2. To increase image resolution
3. To translate speech
4. To reduce embedding dimensions

Correct Answer: 1

Explanation

Access filters help ensure that users retrieve only documents they are authorized to access. In enterprise RAG systems, a search index may contain information belonging to different departments, customers, or security groups. Retrieval should therefore consider the user’s identity and applicable permissions before returning content to the model. Simply hiding sensitive information in the prompt is not sufficient. Developers should enforce authorization at the application or data-access layer and test cases involving users with different permissions to ensure private information cannot be exposed.

Question 332

Which feature can combine keyword and vector retrieval?

1. Hybrid search
2. Speech translation
3. OCR
4. Sentiment analysis

Correct Answer: 1

Explanation

Hybrid search combines traditional keyword retrieval with vector-based semantic retrieval. Keyword search is useful for exact terms, identifiers, and specific phrases, while vector search can find content with related meanings even when the wording differs. Combining both methods can provide broader retrieval coverage for enterprise knowledge applications. Developers should configure ranking and retrieval parameters based on application requirements and evaluate the results using realistic queries. Hybrid search is especially useful when users may search for both exact product names and natural-language descriptions.

Question 333

What is the purpose of Azure Key Vault in an AI solution?

1. Store and manage secrets securely
2. Generate image captions
3. Perform semantic ranking
4. Convert text to speech

Correct Answer: 1

Explanation

Azure Key Vault provides a secure place to manage secrets, keys, and certificates used by applications. AI solutions may need credentials or cryptographic material to access external services, APIs, or other resources. Storing sensitive values in source code or plain configuration files can increase security risks. Developers can use Key Vault together with appropriate identities and access controls to reduce unnecessary exposure of secrets. Permissions should follow least-privilege principles, and applications should avoid logging secret values during normal operation or error handling.

Question 334

Which practice helps determine whether an AI system performs well on unusual inputs?

1. Edge-case testing
2. Increasing password length
3. Removing evaluation data
4. Disabling monitoring

Correct Answer: 1

Explanation

Edge-case testing evaluates how an AI system behaves with unusual, difficult, incomplete, ambiguous, or unexpected inputs. Normal test cases may not reveal failures that occur when users provide malformed requests, very long content, rare terminology, or unexpected combinations of information. Including edge cases in evaluation datasets can reveal weaknesses in prompts, retrieval, model behavior, and output validation. Developers should combine edge-case testing with representative normal cases so that evaluation reflects both everyday usage and situations that could expose important application weaknesses.

Question 335

Which service capability can determine whether text expresses positive or negative sentiment?

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

Correct Answer: 2

Explanation

Azure AI Language provides language-processing capabilities that can analyze text sentiment. Sentiment analysis can help applications identify whether feedback or other text expresses positive, negative, or neutral opinions, depending on the supported analysis capability. It can be useful for customer feedback, reviews, surveys, and support messages. Developers should remember that sentiment analysis is an automated interpretation and can be affected by sarcasm, context, mixed opinions, and domain-specific language. Important business decisions should therefore use appropriate validation and evaluation rather than relying on sentiment alone.

Question 336

What is the purpose of query rewriting in a retrieval system?

1. Improve the search query before retrieval
2. Encrypt stored documents
3. Generate speech audio
4. Assign administrator permissions

Correct Answer: 1

Explanation

Query rewriting transforms or expands a user’s original query into a form that can improve retrieval. This can be useful when the original question is vague, conversational, or uses terminology that differs from the indexed content. A rewritten query may make important concepts more explicit before the search operation occurs. Developers should ensure that rewriting does not introduce incorrect assumptions or remove important constraints. Query rewriting should be evaluated with representative questions because changes to the query can affect both retrieval relevance and downstream generated answers.

Question 337

Which approach helps reduce hallucinations in a knowledge-based AI application?

1. Ground responses in retrieved, relevant sources
2. Increase temperature for every request
3. Remove all source documents
4. Disable retrieval

Correct Answer: 1

Explanation

Grounding responses in relevant retrieved sources can help reduce unsupported statements by providing the model with information related to the user’s request. In a RAG system, retrieval supplies context from approved knowledge sources before generation. The application can also require source references, validate retrieved content, and evaluate groundedness. Grounding does not guarantee that every generated statement is correct, so evaluation and output validation remain important. Developers should also ensure that retrieved sources are current, relevant, and authorized for the requesting user.

Question 338

What should an application do when a temporary service request fails?

1. Use appropriate retry logic
2. Delete the user’s account
3. Disable all monitoring
4. Increase model temperature

Correct Answer: 1

Explanation

Appropriate retry logic can help an application recover from temporary failures such as transient service errors or temporary connectivity issues. Retries should normally use limits and delays, such as exponential backoff, to avoid creating excessive traffic during an outage. Not every error should be retried; permanent validation or authorization failures generally require a different response. Developers should log useful diagnostic information without exposing sensitive data and should monitor repeated failures. Retry policies should be designed according to the behavior and requirements of the dependent service.

Question 339

Why is source citation useful in a RAG response?

1. It helps users identify supporting information
2. It increases audio quality
3. It removes authorization requirements
4. It changes the embedding model

Correct Answer: 1

Explanation

Source citations can help users identify which retrieved documents or passages support an AI-generated response. This can improve transparency and make it easier to verify important information. Citations are particularly useful in enterprise knowledge applications where users may need to review policies, reports, or source documents before acting on an answer. Developers should ensure citations actually correspond to the retrieved content and do not expose documents the user is not authorized to access. Citation support should be evaluated along with retrieval relevance and response groundedness.

Question 340

Which consideration is important when selecting a model for production?

1. Only its name
2. Quality, cost, latency, and workload requirements
3. Only its logo
4. Only its training date

Correct Answer: 2

Explanation

Model selection should consider the application’s actual workload and requirements rather than relying on the model name alone. Important factors can include response quality, supported capabilities, latency, cost, throughput, context requirements, safety characteristics, and deployment availability. A larger or more capable model may not always be necessary for simple tasks, while demanding workloads may require stronger capabilities. Developers should evaluate candidate models using representative datasets and production-like scenarios. Monitoring after deployment is also important because real-world behavior can differ from controlled testing.