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

 

Question 121

Which Azure service is designed to analyze the layout and content of documents?

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

Correct Answer: 2

Explanation

Azure AI Document Intelligence is designed to analyze documents and extract useful information from their content and layout. It can process documents such as invoices, receipts, forms, and other structured or semi-structured files. Depending on the model used, it can identify text, tables, fields, and document structure. This makes it useful for automating manual data-entry tasks and document-processing workflows. Azure AI Language focuses on text analysis, while Azure AI Speech handles spoken language and audio-related scenarios.

Question 122

What does a vector index store for similarity-based search?

1. Numerical vector representations
2. Audio recordings only
3. User passwords
4. Image file names only

Correct Answer: 1

Explanation

A vector index stores numerical representations of content, commonly called embeddings, so that applications can perform similarity-based searches. When a user submits a query, the query can also be converted into a vector and compared with vectors stored in the index. Results with greater semantic similarity can then be retrieved. Vector indexes are especially useful for RAG applications, recommendations, and semantic retrieval. The quality of vector search depends on the embedding model, indexed content, chunking strategy, and search configuration.

Question 123

Which technique can help an agent remember information across multiple interactions?

1. OCR
2. Object detection
3. State or memory management
4. Speech synthesis

Correct Answer: 3

Explanation

State or memory management allows an AI agent to maintain relevant information across interactions. Depending on the application, this may include conversation history, user preferences, task progress, or information retrieved from an external data store. Memory should be designed carefully because storing every piece of conversation data may increase cost and create privacy concerns. Developers should determine what information is actually needed, how long it should be retained, and who can access it. Appropriate security and data-retention controls should accompany persistent agent memory.

Question 124

Which capability is most suitable for detecting whether a customer message is positive or negative?

1. OCR
2. Sentiment analysis
3. Object detection
4. Speech synthesis

Correct Answer: 2

Explanation

Sentiment analysis evaluates written language to determine the expressed sentiment of a message. Common categories include positive, negative, and neutral, depending on the service and configuration. Businesses can use sentiment analysis to process customer reviews, support messages, surveys, and feedback at scale. It can help identify trends and prioritize messages for additional review. However, sentiment models may misunderstand sarcasm, context, or domain-specific language, so developers should evaluate performance using representative examples from the intended application.

Question 125

What is the purpose of semantic ranking in a search solution?

1. Improve the ordering of results based on meaning
2. Convert speech into text
3. Encrypt search documents
4. Generate images

Correct Answer: 1

Explanation

Semantic ranking improves search results by considering the meaning and context of the query and retrieved documents. Traditional keyword matching can return results that contain the right words but do not fully address the user’s intent. Semantic ranking can help place more meaningful results higher in the result set. This is useful in enterprise search and RAG applications where the quality of retrieved context affects the final answer. Developers should evaluate ranking performance using realistic queries and representative content.

Question 126

Which method is useful when an application needs to process thousands of documents without immediate user interaction?

1. Real-time voice conversation
2. Batch processing
3. Manual data entry
4. Interactive chat only

Correct Answer: 2

Explanation

Batch processing is useful when large numbers of documents or records need to be processed without requiring an immediate response to an individual user. Applications can submit work in groups and process it asynchronously according to available resources and service limits. This approach can be appropriate for document extraction, classification, indexing, and other large-scale workloads. Compared with synchronous processing, batch workflows can be designed to handle long-running jobs more efficiently. Developers should also implement monitoring, retries, and appropriate failure handling.

Question 127

Which feature helps an application continue operating when one AI model or endpoint becomes unavailable?

1. Model fallback strategy
2. OCR
3. Sentiment analysis
4. Image resizing

Correct Answer: 1

Explanation

A model fallback strategy allows an application to use an alternative model or endpoint when the primary option cannot process a request. Fallbacks can improve resilience when a service experiences temporary availability problems, capacity limitations, or other failures. The alternative model should be tested to ensure that it meets the application’s minimum quality and capability requirements. Developers should also monitor fallback usage because frequent fallback events may indicate a capacity or reliability problem that requires further investigation.

Question 128

An AI application needs to answer questions from documents while showing where the information came from. What should it provide?

1. Audio effects
2. Image labels
3. Source citations or references
4. Random responses

Correct Answer: 3

Explanation

Source citations or references can show users which documents or sections support an AI-generated response. In a RAG application, the system can retain metadata about retrieved documents and use that information to present relevant sources alongside the generated answer. This can improve transparency and make it easier for users to verify important information. Developers should ensure that citations actually correspond to the retrieved evidence and do not falsely imply support. Access controls must also be applied so users only receive sources they are authorized to view.

Question 129

What is the main purpose of temperature in a generative model request?

1. Control output randomness
2. Control user permissions
3. Control document storage
4. Control network bandwidth

Correct Answer: 1

Explanation

Temperature is commonly used to influence the randomness of generated responses. Lower temperature settings generally encourage more predictable outputs, while higher values can produce greater variation. The appropriate setting depends on the application. A structured business workflow may require consistent responses, while a creative application may benefit from more variation. Temperature does not determine whether a user has permission to access data or control the amount of storage available. Developers should test generation settings with representative prompts and evaluate their effect on application quality.

Question 130

Which capability can extract tables and fields from supported business documents?

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

Correct Answer: 1

Explanation

Azure AI Document Intelligence can extract information from supported business documents, including text, tables, key-value fields, and other structural elements. This capability can automate workflows involving invoices, receipts, forms, and similar documents. Depending on the scenario, developers can use prebuilt models or other supported document-analysis models. Extracted information can then be passed to business applications for validation and processing. Developers should test extraction accuracy with documents that represent the actual layouts, quality levels, and formats expected in production.

Question 131

Why should developers use access-controlled knowledge sources in enterprise RAG systems?

1. To ensure users receive only authorized information
2. To increase image quality
3. To disable search
4. To remove document indexing

Correct Answer: 1

Explanation

Access-controlled knowledge sources help ensure that retrieval results respect the permissions associated with users and documents. In an enterprise RAG system, a search query may match documents containing confidential information that should not be visible to every user. The application should apply authorization rules before providing retrieved content to the model or user. This can involve identity checks, role-based permissions, document metadata, and security trimming. Without appropriate access controls, an AI assistant could unintentionally expose sensitive organizational information.

Question 132

What is a key benefit of using asynchronous processing for long-running AI tasks?

1. It avoids requiring the user to wait for completion
2. It removes all processing costs
3. It guarantees perfect model output
4. It eliminates authentication

Correct Answer: 1

Explanation

Asynchronous processing allows an application to submit a long-running task and continue performing other work while the operation completes. Instead of keeping a user request open for an extended period, the application can track the job and notify the user when processing is finished. This is useful for large document analysis, batch workloads, and other operations that may take significant time. Developers should design appropriate job-status tracking, retry handling, timeout management, and error reporting for asynchronous workflows.

Question 133

Which capability allows a model to receive an image as part of its input?

1. Text-only generation
2. Multimodal input
3. Keyword indexing
4. Speech synthesis

Correct Answer: 2

Explanation

Multimodal input allows a supported AI model to process more than one type of information, such as text combined with an image. For example, an application could send an image of a product along with a question asking the model to identify visible characteristics. Multimodal models can support scenarios such as visual question answering, image analysis, and document interpretation. Developers should verify the supported image formats, size limitations, model capabilities, and cost before implementing multimodal features in production.

Question 134

Which practice can help detect changes in AI application performance after deployment?

1. Monitoring and evaluation
2. Removing telemetry
3. Disabling logs
4. Deleting test data

Correct Answer: 1

Explanation

Monitoring and ongoing evaluation help identify changes in an AI application’s behavior after deployment. Production traffic can differ from development test data, and changes to prompts, models, retrieval indexes, or application code can affect results. Developers can track metrics such as latency, errors, usage, retrieval quality, and other relevant performance indicators. Periodic evaluation using representative datasets can also reveal quality degradation. Monitoring should be implemented with privacy considerations so that sensitive user information is not unnecessarily captured in telemetry.

Question 135

What should an application do if a model returns an invalid value for a required field?

1. Accept it without checking
2. Validate the output and handle the error
3. Give the model administrator access
4. Delete the entire application

Correct Answer: 2

Explanation

The application should validate model-generated values before using them in downstream processing. If a required field is missing, malformed, or outside an expected range, the application can reject the response, request a corrected response, or use a defined fallback process. Validation is particularly important when model output is passed to databases, APIs, financial systems, or other automated workflows. Generative models can produce unexpected results even when instructions are clear. Output validation provides an additional reliability and safety layer.

Question 136

Which approach can reduce unnecessary model costs for simple repetitive tasks?

1. Always use the largest model
2. Select a suitable smaller model
3. Increase prompt length indefinitely
4. Disable caching and reuse

Correct Answer: 2

Explanation

A suitable smaller model can reduce costs and latency when the task does not require the advanced capabilities of a larger model. Examples may include simple classification, extraction, routing, or straightforward text-generation tasks. However, developers should not select a model based only on price. They should evaluate quality, context requirements, throughput, safety, and reliability against representative workloads. Model selection is an application-design decision, and testing can determine whether a smaller model meets the required performance.

Question 137

What is the purpose of a system instruction in an AI agent?

1. Define the agent’s role and behavior
2. Store audio recordings
3. Index images
4. Manage network cables

Correct Answer: 1

Explanation

System instructions provide high-level guidance that defines an AI agent’s intended role, behavior, constraints, and objectives. For example, an agent can be instructed to act as a customer-support assistant, use approved tools, and avoid revealing confidential information. System instructions can improve consistency across interactions, but they should not be treated as a replacement for technical security controls. Developers should combine instructions with authorization, tool restrictions, input validation, output validation, and monitoring to create a more reliable agent.

Question 138

Which capability is used to identify people, organizations, or locations in text?

1. Text-to-speech
2. Named entity recognition
3. Image segmentation
4. OCR

Correct Answer: 2

Explanation

Named entity recognition identifies entities within text and classifies them into recognized categories. Common examples include people, organizations, locations, dates, and other entity types supported by the selected language service. It can be useful for extracting structured information from emails, reports, customer messages, and other unstructured text. For example, an application could identify an organization name from a support request and use that information in a downstream workflow. NER is a natural language processing capability rather than an image-processing task.

Question 139

What is the main purpose of a vector similarity search?

1. Find content with similar semantic representations
2. Convert audio to text
3. Extract tables from invoices
4. Detect faces in photographs

Correct Answer: 1

Explanation

Vector similarity search compares numerical vector representations to identify content that is semantically similar to a query. A query is converted into an embedding and compared with vectors stored in a search index. Documents with similar representations can then be returned as relevant results. This approach is useful when users express an idea using words that differ from those used in the source documents. Vector search is commonly used in RAG systems, recommendation solutions, and semantic information-retrieval applications.

Question 140

Why should an AI solution be tested with edge cases?

1. To identify failures outside normal scenarios
2. To increase the number of passwords
3. To remove all model safeguards
4. To avoid monitoring

Correct Answer: 1

Explanation

Edge-case testing helps developers discover how an AI solution behaves with unusual, ambiguous, incomplete, or unexpected inputs. Real users may provide misspelled queries, very long documents, unsupported formats, unusual language, conflicting instructions, or missing information. Testing these cases can reveal weaknesses in prompts, retrieval, validation, error handling, and model behavior. Developers can then improve the application before deployment. A strong evaluation process should include both common scenarios and realistic edge cases that could affect reliability, safety, or user experience.