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

 

Question 221

Which component stores vector representations for similarity-based retrieval?

1. Speech service
2. Vector index
3. Content filter
4. Secret store

Correct Answer: 2

Explanation

A vector index stores numerical representations, called embeddings, so that applications can perform similarity-based searches. When a user submits a query, the query can also be converted into an embedding and compared with vectors stored in the index. This allows the system to find content with similar meaning rather than relying only on exact keywords. Vector indexes are commonly used in RAG applications and semantic retrieval systems. Developers should select appropriate dimensions, distance metrics, and indexing configurations for their chosen embedding model and workload.

Question 222

Which service can translate spoken language into another language?

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

Correct Answer: 1

Explanation

Azure AI Speech provides capabilities for processing spoken audio, including speech translation scenarios. An application can capture spoken input, recognize the speech, and translate it into another supported language. This can support multilingual meetings, customer-service applications, accessibility solutions, and communication tools. Developers should consider language support, audio quality, speaker conditions, latency, and expected vocabulary when designing the solution. Testing with realistic recordings is important because background noise, accents, overlapping speech, and microphone quality can affect recognition and translation results.

Question 223

What is the main purpose of metadata filtering in a RAG system?

1. Generate speech from text
2. Increase model temperature
3. Restrict retrieval to relevant records
4. Detect objects in images

Correct Answer: 3

Explanation

Metadata filtering restricts search results according to attributes associated with indexed content. For example, documents can contain metadata identifying department, language, date, product, or access level. A search query can then retrieve only documents matching the required conditions. This can improve relevance and help enforce information-access requirements when properly configured. Metadata filtering can be combined with keyword, semantic, vector, or hybrid search. Developers should verify that filters are correctly applied and cannot be bypassed through alternative queries or retrieval paths.

Question 224

Why is prompt injection a concern for AI agents?

1. It increases audio quality
2. Untrusted instructions may influence model behavior
3. It automatically improves retrieval
4. It reduces storage requirements

Correct Answer: 2

Explanation

Prompt injection occurs when untrusted content attempts to influence an AI model’s instructions or behavior. In an agent application, this could cause the model to ignore intended instructions, reveal sensitive information, or attempt inappropriate tool actions. Retrieved documents, web pages, emails, and user inputs can all contain untrusted content. Developers should separate trusted instructions from untrusted data, restrict tool permissions, validate tool arguments, and apply authorization controls. Sensitive operations may also require human approval to reduce the impact of successful injection attempts.

Question 225

Which capability helps an AI application call an external function or API?

1. Tool or function calling
2. OCR
3. Sentiment analysis
4. Image classification

Correct Answer: 1

Explanation

Tool or function calling allows a generative AI model to request an external function or API when an application determines that a tool is needed. For example, an agent could call an order-status API to retrieve information that is not contained in its model knowledge. The application remains responsible for executing the tool and should validate the generated arguments before doing so. Authentication, authorization, input validation, and error handling should also be applied. Tool descriptions should clearly communicate the purpose and expected inputs of each available function.

Question 226

What should be included when evaluating an AI application before production?

1. Only successful examples
2. Representative and challenging test cases
3. Only empty inputs
4. Only administrator requests

Correct Answer: 2

Explanation

AI evaluation should include representative examples that reflect actual production use as well as challenging cases and important edge conditions. Testing only successful or simple inputs can hide weaknesses in retrieval, reasoning, safety, structured output, or error handling. An evaluation dataset should reflect the application’s target users, languages, document types, and common workflows where appropriate. Developers can compare results across different prompts, models, and configurations using the same evaluation data. Continuous evaluation is also useful because application behavior can change after updates.

Question 227

Which service is designed to extract fields and tables from supported business documents?

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

Correct Answer: 3

Explanation

Azure AI Document Intelligence is designed to analyze documents and extract information from their content and structure. Depending on the selected capability, it can identify text, fields, tables, and other elements in supported documents. This makes it useful for invoices, receipts, forms, applications, and similar business records. Extracted information can then be validated and integrated into downstream systems. Developers should test the selected model with real document layouts because variations in formatting, scan quality, and document structure can affect extraction results.

Question 228

What is a benefit of semantic ranking in search?

1. It can improve the ordering of relevant results
2. It converts text to speech
3. It stores API credentials
4. It detects spoken languages

Correct Answer: 1

Explanation

Semantic ranking can improve the ordering of search results by considering the meaning and context of the query and retrieved content. This can help place more relevant documents above results that may share keywords but provide less useful information. In AI applications, better ranking can provide a language model with higher-quality context and potentially reduce irrelevant information. Developers should still evaluate semantic ranking using representative queries because relevance depends on the content, search configuration, and specific requirements of the application.

Question 229

Which approach is most appropriate for frequently changing company policies?

1. Retrain the model after every policy change
2. Remove the policies from the knowledge base
3. Retrieve the current policies from an updated knowledge source
4. Store policies only in the system prompt

Correct Answer: 3

Explanation

Frequently changing company policies are generally better maintained in an external knowledge source that can be updated and retrieved when needed. A RAG solution can provide the latest approved policy information to a generative model without requiring model retraining for every change. The knowledge source should have an appropriate update and indexing process, along with access controls where necessary. Developers should also evaluate retrieval freshness and ensure outdated documents are removed, replaced, or clearly versioned so that the model does not use obsolete information.

Question 230

Which parameter generally controls randomness in generated model responses?

1. Temperature
2. Vector dimension
3. Chunk overlap
4. Token identifier

Correct Answer: 1

Explanation

Temperature is a model-generation parameter that generally influences the randomness of generated responses. Lower values can make outputs more consistent and predictable, while higher values can produce greater variation. The appropriate setting depends on the application. Tasks requiring consistent structured responses may benefit from lower randomness, while creative applications may allow more variation. Temperature does not replace output validation or guarantee factual accuracy. Developers should test the setting with representative prompts and evaluate whether changes improve the desired behavior without introducing unnecessary variability.

Question 231

Why should an AI application monitor token usage?

1. To detect image defects
2. To track consumption and manage costs
3. To identify microphone hardware
4. To replace authorization

Correct Answer: 2

Explanation

Monitoring token usage helps developers understand how much model input and output an application consumes. Token consumption can affect service costs, latency, and throughput, particularly when prompts contain large amounts of retrieved context. Monitoring can also reveal unusual usage patterns that may indicate inefficient prompts, excessive retrieval, or unexpected traffic. Developers can use this information to optimize prompts, limit unnecessary context, select suitable models, and establish usage controls. Token monitoring should be considered alongside other operational metrics such as latency, errors, and request volume.

Question 232

What does a model deployment provide to an application?

1. An accessible configured model endpoint or resource
2. A physical storage device
3. A document scanner
4. A user password database

Correct Answer: 1

Explanation

A model deployment makes a selected model available for application use through a configured service resource or endpoint. Depending on the platform and model, deployment configuration can include the model version, deployment name, capacity, and other relevant settings. Applications can then send requests to the appropriate deployment. Developers should monitor deployment performance, capacity, cost, and availability. They should also ensure that applications use the intended model and configuration, especially when multiple deployments exist for different workloads or environments.

Question 233

Which technique helps retrieve documents based on concepts rather than exact words?

1. File extension filtering
2. Keyword-only search
3. Vector search
4. Audio transcription

Correct Answer: 3

Explanation

Vector search uses embeddings to identify content based on semantic similarity. Because embeddings represent aspects of meaning, a query can retrieve relevant documents even when they use different wording from the user’s original question. This is useful for knowledge bases, support systems, and RAG applications. Vector search quality depends on the embedding model, document chunking, index configuration, and source content. Developers should test retrieval with representative queries and compare results against expected relevant documents to determine whether the selected configuration provides adequate search quality.

Question 234

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

1. Define expected behavior and boundaries
2. Store vector embeddings
3. Process scanned invoices
4. Detect faces in images

Correct Answer: 1

Explanation

A system instruction defines high-level behavior, role, constraints, and response expectations for an AI agent. For example, it can instruct an agent to answer using approved knowledge sources, follow a particular communication style, or request clarification when required information is missing. System instructions can improve consistency but should not be considered a complete security mechanism. Applications should independently enforce authorization, tool restrictions, data-access rules, and validation. Clear instructions are especially useful when an agent has multiple tools or operates across several business workflows.

Question 235

Which feature can help identify the overall emotional tone of customer feedback?

1. OCR
2. Sentiment analysis
3. Vector indexing
4. Speech synthesis

Correct Answer: 2

Explanation

Sentiment analysis examines text to determine its expressed sentiment or emotional polarity according to the supported service capabilities. It can help organizations analyze customer feedback, reviews, surveys, support messages, and other textual content. Results can provide signals about whether content is positive, negative, neutral, or otherwise classified by the service. Sentiment analysis should not be treated as a perfect representation of a person’s actual feelings because language can contain sarcasm, ambiguity, or domain-specific expressions. Developers should evaluate results using representative data.

Question 236

Why can asynchronous processing be useful for large document workloads?

1. It allows long-running work to be handled without blocking the main request
2. It disables all authentication
3. It guarantees perfect extraction
4. It converts documents into audio

Correct Answer: 1

Explanation

Asynchronous processing is useful when a task may take significant time to complete. Instead of keeping a user request waiting for the entire operation, an application can submit the workload and process the result when it becomes available. This approach can improve responsiveness and scalability for workloads such as large document analysis, batch processing, or lengthy AI operations. Developers should design appropriate status tracking, retry handling, timeout behavior, and failure reporting. The exact implementation depends on the service and workload requirements.

Question 237

Which control limits what an authenticated user or service can access?

1. Authentication
2. OCR
3. Authorization
4. Embedding generation

Correct Answer: 3

Explanation

Authorization determines which resources or actions an authenticated identity is permitted to access. Authentication establishes who or what the identity is, while authorization determines what that identity can do. In AI applications, authorization is important when agents retrieve private documents, call business APIs, or perform sensitive operations. Developers should apply appropriate roles and permissions and follow least-privilege principles. Authorization should be enforced by the application or underlying service rather than relying solely on model instructions to prevent unauthorized access.

Question 238

What is a useful reason to use hybrid search in enterprise RAG?

1. It combines lexical and semantic retrieval signals
2. It replaces document storage
3. It converts audio into images
4. It eliminates access controls

Correct Answer: 1

Explanation

Hybrid search combines lexical or keyword-based retrieval with semantic or vector-based retrieval. This is useful because different queries benefit from different search signals. Exact identifiers, product codes, and specialized terms may perform well with keyword matching, while natural-language questions may benefit from semantic similarity. Combining these approaches can improve the set of candidate documents supplied to downstream ranking or generation components. Developers should test hybrid search with representative enterprise queries and tune the configuration based on actual retrieval performance.

Question 239

Which practice helps protect an AI application from excessive request traffic?

1. Removing request limits
2. Disabling monitoring
3. Rate limiting
4. Increasing document size

Correct Answer: 3

Explanation

Rate limiting restricts the number or frequency of requests that an application or service accepts during a specified period. It can help protect systems from excessive traffic, unexpected spikes, resource exhaustion, and certain abusive usage patterns. Rate limiting can also help control costs and prevent service quotas from being consumed too quickly. Developers should choose limits based on expected workloads and service capabilities. Applications should also handle limit responses appropriately, using controlled retry behavior where suitable rather than immediately generating additional requests.

Question 240

What should an application do when a model returns invalid structured data?

1. Accept it without checking
2. Validate the response and handle the failure
3. Give the model administrator access
4. Disable all logging

Correct Answer: 2

Explanation

When a model produces structured data that does not meet the required schema, the application should validate the response and handle the failure safely. Depending on the workload, it may retry with appropriate instructions, request another generation, use a fallback process, or return an error. Blindly accepting invalid data can cause downstream systems to behave incorrectly. Validation should check both structure and important business rules. Developers should monitor validation failures because repeated failures can indicate problems with prompts, model selection, schemas, or application logic.