View Full Microsoft AI-102 Exam Dumps and Practice Test Dumps.

 

Question 381

Which Microsoft Foundry capability helps evaluate an AI application using predefined datasets and quality criteria?

1. Model evaluation
2. OCR
3. Image analysis
4. Speech synthesis

Correct Answer: 1

Explanation

Model evaluation provides a structured approach for measuring AI application performance against datasets and defined evaluation criteria. Developers can use representative inputs to assess dimensions such as relevance, groundedness, coherence, fluency, safety, similarity, or task success, depending on the scenario. Evaluation is useful when comparing models, prompts, retrieval strategies, or application versions because it provides measurable evidence instead of relying only on a few manually reviewed responses. Teams should maintain suitable evaluation datasets and repeat tests after important changes. This helps identify regressions and confirms whether modifications actually improve the application’s intended behavior.

Question 382

An application needs to find documents that contain the exact term “INV-4582.” Which retrieval method is most appropriate?

1. Image captioning
2. Keyword search
3. Pronunciation assessment
4. Speech translation

Correct Answer: 2

Explanation

Keyword search is appropriate when an application needs to locate an exact identifier or specific textual term such as an invoice number. Lexical retrieval is particularly useful for product codes, document identifiers, names, and other values where exact or close textual matching is important. Vector search can be useful for conceptual queries, but embeddings may not provide the same precision for unique identifiers. Developers can combine keyword retrieval with vector search when an application needs both exact matching and semantic similarity. Search behavior should be tested using representative identifiers, spelling variations, and realistic user queries.

Question 383

Which Azure AI Search technique can retrieve semantically similar documents even when they do not contain the same words as the query?

1. Faceting
2. Vector search
3. Sorting
4. Field mapping

Correct Answer: 2

Explanation

Vector search retrieves documents according to the similarity between embeddings rather than depending entirely on matching words. This allows an application to identify content that expresses a similar concept using different terminology. For example, a user asking about reducing automobile fuel consumption could retrieve documents discussing vehicle efficiency even if the exact query wording is absent. Developers must generate embeddings consistently and configure the vector field with compatible dimensions and search settings. Retrieval quality should be evaluated with representative queries because chunking, embedding models, metadata filters, and ranking configuration can all influence which documents are returned.

Question 384

A developer wants to prevent an AI agent from accessing resources beyond its assigned business function. Which security principle should be applied?

1. Unlimited access
2. Shared credentials
3. Least privilege
4. Anonymous access

Correct Answer: 3

Explanation

The principle of least privilege limits an AI agent’s permissions to only those resources and operations required for its authorized business function. This reduces the potential impact if the agent makes an incorrect tool selection or receives manipulated instructions. For example, an agent that only needs to read customer records should not automatically receive permission to delete records or modify account settings. Developers should combine least privilege with authentication, authorization, parameter validation, monitoring, and approval controls for sensitive actions. Permissions should be reviewed regularly because application requirements can change and unnecessary access may accumulate over time.

Question 385

Which Document Intelligence model is intended to extract commonly supported fields from invoices without creating a custom extraction model?

1. Prebuilt invoice model
2. Custom NER model
3. Speech model
4. Image classifier

Correct Answer: 1

Explanation

The prebuilt invoice model is designed to extract commonly required invoice information without requiring developers to train a custom extraction model from scratch. Depending on supported capabilities, extracted information can include fields such as vendor details, invoice dates, totals, and line-item information. This can significantly simplify document-processing solutions when the application’s requirements match the model’s supported fields. Developers should validate the extracted values because document layouts and scan quality can vary. If an organization requires specialized fields or unusual document structures that the prebuilt model does not adequately support, a custom Document Intelligence model may be more appropriate.

Question 386

Which Azure AI Speech feature controls how a synthesized voice pronounces specialized terms?

1. SSML
2. Vector indexing
3. Semantic ranking
4. Document classification

Correct Answer: 1

Explanation

SSML can provide additional control over synthesized speech, including pronunciation-related behavior and other speech characteristics supported by the selected voice. This is useful when an application contains specialized terminology, names, abbreviations, or other words that require controlled pronunciation. Developers can incorporate appropriate SSML elements into the text sent to the speech synthesis service and test the result with the selected voice. SSML is applied during text-to-speech processing and does not perform speech recognition. Teams should verify supported SSML features because behavior can vary depending on the speech service configuration and selected voice.

Question 387

An AI Search index must support filtering documents by a field named Department. Which field capability should be enabled?

1. Retrievable
2. Filterable
3. Vectorized
4. Faceted only

Correct Answer: 2

Explanation

The Department field should be configured as filterable when the application needs to restrict search results according to department values. For example, a query can request documents where Department equals Finance or Engineering. Filterable fields can be combined with keyword, vector, semantic, or hybrid retrieval. Developers should ensure that the field contains consistent metadata and uses an appropriate data type. Filtering should not be confused with authorization: a user-controlled filter should never be the only mechanism protecting sensitive documents. Access decisions should be enforced by the application and appropriate identity-based controls before returning restricted search results.

Question 388

A customer-support system needs to identify important topics from each incoming message. Which capability should it use?

1. Text-to-speech
2. Key phrase extraction
3. Object detection
4. Document classification

Correct Answer: 2

Explanation

Key phrase extraction identifies important words or phrases that represent significant concepts in a text. It can help customer-support systems identify topics, organize messages, generate metadata, or perform basic content analysis. Unlike custom text classification, it does not require predefined categories such as Billing or Technical Support. Unlike named entity recognition, it is not specifically designed to identify categories such as people, organizations, or locations. Developers should evaluate extracted phrases against representative customer messages because short text, informal language, and domain-specific terminology can affect results. Key phrases should be treated as analytical signals rather than complete summaries.

Question 389

A RAG application needs to provide users with the original source document for generated answers. Which information should the application preserve during indexing?

1. Source metadata
2. Audio volume
3. Image brightness
4. Speech pitch

Correct Answer: 1

Explanation

Source metadata allows retrieved chunks to remain associated with their original documents and locations. Useful metadata can include document identifiers, filenames, URLs, page numbers, section names, or other source references. Preserving this information enables applications to provide citations or links to supporting documents and makes generated answers easier to verify. Metadata can also support filtering and document-level security when properly designed. Developers should preserve source identifiers throughout document ingestion, chunking, embedding, and indexing. When an answer is generated, the application can then associate the selected evidence with the corresponding source information.

Question 390

Which Azure AI Vision capability extracts printed or handwritten text from supported images?

1. OCR
2. Semantic ranking
3. Entity linking
4. Speech recognition

Correct Answer: 1

Explanation

Optical character recognition extracts textual content from supported images and documents. It can be used for photographs, scanned pages, forms, signs, receipts, and other visual material containing text. After extraction, the resulting text can be passed to additional AI services for translation, classification, entity recognition, summarization, or search indexing. Developers should evaluate OCR using representative images because text size, image quality, orientation, handwriting, background noise, and layout can affect recognition accuracy. OCR should not be confused with object detection, which identifies objects and their locations, or image captioning, which generates a natural-language description of the broader image.

Question 391

A developer wants an agent to retrieve current stock prices from an external service instead of relying on model knowledge. What should be used?

1. External tool
2. Static system prompt
3. OCR skill
4. Image caption

Correct Answer: 1

Explanation

An external tool allows an AI agent to access information or functionality that exists outside the model itself. Current stock prices are time-sensitive and therefore should be retrieved from an appropriate authorized service rather than relying on static model knowledge. The tool should define the required inputs and expected outputs clearly so the agent can invoke it appropriately. Developers should authenticate access, validate parameters, restrict permissions, and handle service failures. Applications should also distinguish retrieved real-time data from model-generated explanations. Tool activity should be monitored so unexpected or unauthorized requests can be investigated.

Question 392

Which Azure AI Search feature lets users refine results by selecting values such as “HR” or “Finance”?

1. Vector dimensions
2. Facets
3. SSML
4. OCR

Correct Answer: 2

Explanation

Facets provide grouped counts for values within structured search fields and allow users to refine result sets interactively. For example, an enterprise search interface can display Department values such as HR, Finance, and Sales along with the number of matching documents. Users can then select a value to narrow the results. Facets require suitable indexed fields and are commonly used alongside filters. They do not themselves perform semantic understanding or vector similarity. Developers should choose metadata fields that provide meaningful refinement options and verify that the indexed values are consistent so facet counts accurately represent the available documents.

Question 393

Which evaluation metric is especially important for determining whether a RAG response is supported by retrieved evidence?

1. Groundedness
2. Image resolution
3. Audio duration
4. Search latency only

Correct Answer: 1

Explanation

Groundedness evaluates whether generated content is supported by the information supplied as context. This is a key consideration for RAG applications because retrieval is intended to provide evidence that supports the model’s response. A response can sound fluent and directly address the user’s question while still containing unsupported claims. Developers should therefore measure groundedness separately from relevance and other quality dimensions. Evaluation datasets should include questions that require information from different retrieved sources and cases where the available evidence is insufficient. Monitoring groundedness after model or retrieval changes can help identify unsupported-generation regressions.

Question 394

A team wants to improve a search application that returns conceptually related documents but misses exact product identifiers. What retrieval strategy should they consider?

1. Hybrid search
2. Speech translation
3. Image captioning
4. Text summarization

Correct Answer: 1

Explanation

Hybrid search combines lexical and vector retrieval, making it suitable when an application needs both exact term matching and semantic similarity. Exact product identifiers are often handled effectively by keyword retrieval, while conceptual descriptions can benefit from vector similarity. Combining these approaches can improve coverage across different query types. Developers should evaluate the resulting ranking with representative examples containing identifiers, natural-language descriptions, abbreviations, and mixed queries. The application may also need suitable filters, semantic ranking, or scoring configuration. Hybrid retrieval should be measured using actual search requirements rather than assuming that either keyword or vector search alone is always sufficient.

Question 395

Which Microsoft Foundry feature can help developers compare different prompts using the same test inputs?

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

Correct Answer: 1

Explanation

Prompt evaluation allows developers to assess different prompt versions using consistent test inputs and defined evaluation criteria. Keeping the test data consistent makes it easier to determine whether a prompt change affected application behavior. Developers can compare dimensions such as relevance, groundedness, task success, coherence, or safety according to the application’s requirements. This is especially useful for generative AI applications because small instruction changes can produce different behavior across diverse inputs. Teams should maintain representative evaluation datasets and include edge cases rather than relying on a few manually selected examples. Evaluation results can provide a baseline for future changes.

Question 396

An application must detect whether a user message is English, French, or Urdu before selecting a processing workflow. Which capability should be used?

1. Entity linking
2. Language detection
3. Image analysis
4. Sentiment analysis

Correct Answer: 2

Explanation

Language detection identifies the language represented in a text input and can be used to route content to an appropriate downstream workflow. For example, an application may detect the user’s language before selecting a translation model, language-specific classifier, or supported speech configuration. Developers should consider that very short messages, mixed-language text, unusual spelling, and ambiguous terms can affect detection results. Language detection is different from translation because it identifies the language rather than converting the content into another language. Applications should test the capability using realistic multilingual inputs before depending on the result for important automated routing decisions.

Question 397

Which Azure AI Search method performs an exact nearest-neighbor comparison instead of using an approximate graph?

1. Semantic ranker
2. Exhaustive KNN
3. Hybrid search
4. Keyword search

Correct Answer: 2

Explanation

Exhaustive KNN performs exact nearest-neighbor comparisons between the query vector and available vectors. Unlike approximate methods such as HNSW, it does not rely on a graph-based approximation to reduce the search space. This can provide exact similarity results but may require more computational resources as the number of vectors increases. Developers can use exhaustive KNN when exact retrieval is important or when validating another vector-search configuration. The choice between exact and approximate retrieval should consider dataset size, latency, resource requirements, and application needs. Testing with representative data helps determine the appropriate search strategy.

Question 398

A developer needs to create a category model for messages labeled “Refund,” “Shipping,” and “Payment.” Which capability should be selected?

1. Custom text classification
2. OCR
3. Entity linking
4. Speech-to-text

Correct Answer: 1

Explanation

Custom text classification allows developers to create application-specific categories for text. In this example, messages can be labeled Refund, Shipping, or Payment and used as training examples for the classification model. The examples should represent the variety of language users are likely to submit, including different wording, spelling, and levels of detail. After training, the model should be evaluated with separate test messages rather than only the examples used during training. Custom classification can then support automated routing, queue assignment, workflow selection, or prioritization. Developers should periodically evaluate the model as business terminology and customer behavior change.

Question 399

Which control is most appropriate when an AI agent is about to perform a sensitive operation such as deleting a customer’s account?

1. Human approval
2. Larger context window
3. More embeddings
4. Image captioning

Correct Answer: 1

Explanation

Human approval can provide an additional safeguard before an AI agent performs a sensitive or irreversible operation. Deleting a customer account can have significant consequences, so an application may require the agent to prepare the action while an authorized user confirms it before execution. This should be combined with authentication, authorization, parameter validation, and appropriate logging. Developers should not rely solely on the model’s judgment to determine whether an operation is safe. Approval interfaces should clearly present the requested action and relevant details so the authorized reviewer can understand what will happen before confirming it.

Question 400

A team changes the generative model used by an AI application and wants to determine whether response quality has declined. What should they do?

1. Disable evaluation
2. Compare against a representative evaluation baseline
3. Remove retrieved context
4. Increase all user permissions

Correct Answer: 2

Explanation

A representative evaluation baseline allows the team to compare the application’s behavior before and after changing the generative model. The same test dataset and relevant evaluation criteria should be used where possible so that differences can be attributed more reliably to the model change. Depending on the application, the team may evaluate relevance, groundedness, task success, coherence, fluency, safety, or other appropriate dimensions. Reviewing failed examples is also important because aggregate scores may hide specific weaknesses. Maintaining evaluation baselines helps teams detect regressions and make model changes based on measurable application requirements rather than isolated responses.