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

 

Question 61

Which Azure AI service is designed to extract text and structured fields from invoices?

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

Correct Answer: 4

Explanation

Azure AI Document Intelligence is designed to analyze documents and extract useful information from them. It supports prebuilt models for common documents such as invoices, receipts, and identity documents. It can identify fields, values, tables, and other structured content from document files. This makes it useful when an application needs to automatically process business documents without manually entering their information. Azure AI Speech handles audio, Azure AI Vision focuses on images, and Azure AI Language provides natural language processing capabilities.

Question 62

What is the main purpose of embeddings in an AI application?

1. Representing text as numerical vectors
2. Converting text directly into speech
3. Detecting objects in images
4. Translating audio into another language

Correct Answer: 1

Explanation

Embeddings represent text, images, or other content as numerical vectors that capture meaningful relationships between items. Similar content generally produces vectors that are closer together in vector space. This makes embeddings useful for semantic search, recommendation systems, document retrieval, and retrieval-augmented generation. For example, two questions with different wording but similar meaning can have similar embeddings. The vectors can then be stored in a vector-enabled search system and compared to retrieve relevant information.

Question 63

An application needs to search documents using both keywords and semantic meaning. Which search approach should it use?

1. Keyword-only search
2. Image classification
3. Hybrid search
4. Speech synthesis

Correct Answer: 3

Explanation

Hybrid search combines traditional keyword-based search with vector or semantic search. Keyword search is useful when exact terms, names, codes, or phrases matter, while vector search can identify content with similar meaning even when different words are used. Combining both approaches can improve retrieval quality for AI applications that search large document collections. Azure AI Search supports hybrid search scenarios, making it suitable for retrieval-augmented generation and other applications where both lexical matching and semantic similarity are valuable.

Question 64

What should an AI developer use to prevent an application from returning malformed JSON?

1. Image classification
2. Structured output validation
3. Speech recognition
4. Vector indexing

Correct Answer: 2

Explanation

Structured output validation helps ensure that an AI model’s response follows a required format, such as a JSON schema. This is important when the application passes model-generated data to another component that expects specific fields and data types. Developers can define the expected structure and validate the response before processing it. If the response is invalid, the application can reject it, retry the request, or apply an error-handling process. This improves reliability and reduces failures caused by unexpected model output.

Question 65

Which feature allows a generative AI application to call an external function or API?

1. OCR
2. Speech synthesis
3. Vector embedding
4. Tool or function calling

Correct Answer: 4

Explanation

Tool or function calling allows a generative AI application to interact with external functions, APIs, or services. Instead of generating an answer only from its internal model knowledge, the model can identify when a tool is needed and provide the required arguments. The application then executes the function and can return the result to the model. This approach is useful for tasks such as retrieving current information, checking databases, calculating values, or performing business operations while keeping tool execution under application control.

Question 66

Why is chunking commonly used when preparing documents for RAG?

1. To convert documents into images
2. To translate documents
3. To divide large documents into smaller searchable sections
4. To remove all document metadata

Correct Answer: 3

Explanation

Chunking divides large documents into smaller sections before they are indexed for retrieval. Smaller chunks allow a search system to retrieve the specific portion of a document that is relevant to a user’s question instead of returning an entire large document. Good chunking can improve retrieval accuracy and reduce the amount of unnecessary context sent to a language model. Developers may also use overlap between chunks to preserve important information that appears near the boundary between two sections.

Question 67

Which authentication method can help an Azure-hosted application access resources without storing passwords in code?

1. Managed identity
2. Plain-text credentials
3. Hard-coded API keys
4. Embedded passwords

Correct Answer: 1

Explanation

Managed identity allows an Azure resource to authenticate to supported services without requiring developers to store credentials directly in application code. Azure can manage the identity and provide authentication tokens when the application accesses authorized resources. This reduces the need to maintain passwords or API keys in source code. Developers can then use Azure role-based access control to grant only the permissions required by the application. Managed identities are especially useful for production applications that need secure access to Azure services.

Question 68

What is the primary purpose of a reranker in an AI search solution?

1. Generate speech from text
2. Reorder retrieved results based on relevance
3. Extract text from images
4. Create user authentication tokens

Correct Answer: 2

Explanation

A reranker evaluates an initial set of search results and reorders them according to their relevance to the user’s query. An initial retrieval stage may quickly identify many potentially useful documents, but some results may be more relevant than others. Reranking helps place the most useful results near the top. In retrieval-augmented generation systems, this can improve the quality of the context provided to the language model. Better-ranked context can help the model produce more relevant and grounded responses.

Question 69

A company wants an AI application to answer questions using its frequently updated internal documents. What should it use?

1. Model fine-tuning only
2. Speech translation
3. Retrieval-augmented generation
4. Image segmentation

Correct Answer: 3

Explanation

Retrieval-augmented generation allows an application to retrieve relevant information from an external knowledge source before generating a response. This is useful for company documents that change frequently because the underlying language model does not need to be retrained every time information changes. Documents can be updated in the retrieval system, and future queries can use the latest indexed content. This approach also allows organizations to keep knowledge in controlled repositories while providing relevant context to the generative model.

Question 70

Which Azure AI capability can convert spoken audio into written text?

1. Text-to-speech
2. Optical character recognition
3. Sentiment analysis
4. Speech-to-text

Correct Answer: 4

Explanation

Speech-to-text converts spoken language from an audio source into written text. Azure AI Speech provides speech recognition capabilities that can process microphone input, recorded audio, and other supported sources. Speech-to-text is commonly used for meeting transcription, voice-controlled applications, call-center analysis, and accessibility features. It is different from text-to-speech, which performs the opposite operation by generating spoken audio from text. Optical character recognition focuses on extracting written or printed text from images and documents rather than audio.

Question 71

Which principle limits an AI application’s permissions to only what it needs?

1. Least privilege
2. Data duplication
3. Maximum access
4. Public access

Correct Answer: 1

Explanation

The principle of least privilege means that an application, user, or service should receive only the permissions required to perform its intended tasks. For AI applications, this can reduce the potential impact if credentials are compromised or a component behaves unexpectedly. Developers can apply role-based access control and managed identities to implement this principle in Azure environments. For example, an application that only needs to read documents should not automatically receive permission to modify or delete those documents.

Question 72

What is the purpose of a system prompt or system instruction?

1. Store vector embeddings
2. Define the model’s behavior and high-level rules
3. Convert speech to text
4. Detect objects in images

Correct Answer: 2

Explanation

System instructions define important behavior, constraints, and goals for a generative AI model. They can specify the assistant’s role, response style, rules, limitations, and other high-level requirements. For example, an enterprise assistant can be instructed to answer using retrieved company information and avoid revealing confidential data. System instructions are different from user prompts because they establish application-level behavior. Developers should still combine instructions with access controls, validation, and other security measures because prompts alone cannot provide complete security.

Question 73

Which capability is most appropriate for identifying named people, organizations, and locations in text?

1. Speech synthesis
2. Image generation
3. Object detection
4. Named entity recognition

Correct Answer: 4

Explanation

Named entity recognition identifies specific types of entities within text, such as people, organizations, locations, dates, or other recognized categories. It is a natural language processing task that can help applications organize and analyze unstructured text. For example, a customer-support system could identify a company name and location from incoming messages. Object detection is used for identifying objects in images, while speech synthesis generates audio from text. Named entity recognition focuses specifically on extracting meaningful entities from language.

Question 74

What is the main purpose of grounding a generative AI response?

1. Increase the model’s temperature
2. Convert text into embeddings only
3. Connect responses to relevant factual information
4. Replace authentication

Correct Answer: 3

Explanation

Grounding provides a generative AI model with relevant factual information that it can use when producing an answer. In enterprise applications, grounding is often achieved by retrieving information from approved documents, databases, or search indexes. This can reduce the likelihood of unsupported responses and help the model answer questions using current organizational information. Grounding does not replace authentication or authorization. Access controls are still required to ensure that the application retrieves only information the requesting user or service is allowed to access.

Question 75

A developer needs to extract printed text from an image. Which capability should be considered?

1. Speech-to-text
2. Optical character recognition
3. Sentiment analysis
4. Text-to-speech

Correct Answer: 2

Explanation

Optical character recognition, commonly called OCR, extracts written or printed characters from images or scanned documents. It is useful for digitizing paper documents, reading signs, extracting text from photographs, and processing scanned files. Azure AI Vision and document-focused services provide OCR-related capabilities for different scenarios. Speech-to-text works with spoken audio, while text-to-speech creates audio from written content. Sentiment analysis determines the emotional tone of text and does not extract characters from images.

Question 76

Why should AI applications use representative test data during evaluation?

1. To measure performance on realistic scenarios
2. To increase API credentials
3. To disable model safety features
4. To remove all retrieved documents

Correct Answer: 1

Explanation

Representative test data helps developers evaluate how an AI system performs under conditions similar to real-world usage. A test set should cover the types of inputs, languages, document formats, edge cases, and user scenarios that the application is expected to encounter. Measuring performance only on simple or unrealistic examples can produce misleading results. Evaluation can examine factors such as accuracy, relevance, groundedness, and response quality. Developers can use the results to identify weaknesses and improve prompts, retrieval, models, or application logic.

Question 77

Which Azure service is commonly used to store and search indexed enterprise content for RAG applications?

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

Correct Answer: 4

Explanation

Azure AI Search provides search capabilities that can be used to index and retrieve enterprise content for AI applications. It supports keyword, semantic, vector, and hybrid search scenarios. In a RAG architecture, documents can be processed and indexed, while user queries are used to retrieve relevant content. The retrieved information can then be supplied to a generative model as context. This architecture helps applications answer questions using organizational knowledge without requiring the language model itself to contain every document.

Question 78

What should an application do when an AI service request fails temporarily?

1. Ignore the failure
2. Use appropriate error handling and retry logic
3. Delete the user’s data
4. Disable authentication permanently

Correct Answer: 2

Explanation

AI applications should handle temporary service failures gracefully. Appropriate retry logic can attempt the request again when failures are caused by transient conditions such as temporary network problems or service throttling. Developers should use controlled retries with suitable delays rather than continuously sending requests. Error handling should also provide useful fallback behavior when retries fail. Logging and monitoring can help identify recurring problems. The exact retry strategy should consider the service’s guidance, request type, latency requirements, and application reliability goals.

Question 79

What does temperature generally control in a generative AI model?

1. Authentication permissions
2. Document indexing speed
3. Randomness of generated responses
4. Image resolution

Correct Answer: 3

Explanation

Temperature is a generation parameter that generally influences the randomness of a language model’s output. Lower values tend to produce more predictable and consistent responses, while higher values can produce more varied output. The appropriate setting depends on the application. Tasks requiring consistent structured responses may benefit from lower randomness, while creative generation may use a higher value. Temperature does not control authentication, document indexing, or image resolution. Developers should evaluate the setting using representative application data rather than relying only on assumptions.

Question 80

Which practice helps protect sensitive information in an AI application?

1. Data minimization
2. Sharing all data with every service
3. Storing secrets in source code
4. Granting every user administrator access

Correct Answer: 1

Explanation

Data minimization means collecting, processing, and retaining only the information necessary for the application’s purpose. This reduces the amount of sensitive information exposed to AI models, services, logs, and other components. Developers should also apply appropriate access controls, secure secret storage, encryption, and retention policies. For example, an application should avoid sending unnecessary personal information to a model when only a small portion of the data is required. Data minimization is an important practice for improving privacy and reducing unnecessary exposure.