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

 

Question 341

Which Azure AI Search feature can return highlighted portions of documents that best match a user’s query?

1. Semantic captions
2. Vector profile
3. Data source
4. Skillset

Correct Answer: 1

Explanation

Semantic captions provide concise passages from documents that are particularly relevant to a user’s search query. They can help applications present useful supporting text without requiring users to open every complete document returned by a search. This is especially helpful in knowledge bases and enterprise search experiences where users need quick evidence for why a result is relevant. Semantic captions are generated from indexed content and are associated with semantic search capabilities. Developers should configure appropriate searchable content fields and evaluate returned captions against representative queries to ensure that the displayed excerpts accurately reflect the information users need.

Question 342

A developer needs to map a source field named Customer_ID to an index field named CustomerId during indexing. Which feature should be configured?

1. Semantic ranker
2. Field mapping
3. Scoring profile
4. Vector search

Correct Answer: 2

Explanation

Field mappings allow an Azure AI Search indexer to associate fields from a data source with fields in the target search index when their names or structures do not directly match. For example, a source property called Customer_ID can be mapped to a target field named CustomerId. Field mappings can also support certain transformations during indexing. Developers should verify that the destination field has a compatible type and configuration. Correct mappings are important because a mismatch can result in missing or incorrectly indexed information. After changing mappings, the indexing process should be monitored for errors and validated against expected index content.

Question 343

A company wants an AI application to answer questions using information from newly uploaded internal files. Which architecture is most appropriate?

1. Static model-only generation
2. RAG
3. Image classification
4. Speech synthesis

Correct Answer: 2

Explanation

Retrieval-augmented generation, or RAG, allows an AI application to retrieve relevant information from an external knowledge source before generating a response. For newly uploaded internal files, the documents can be processed, chunked, indexed, and represented using searchable text and embeddings. When a user asks a question, relevant content is retrieved and supplied as context to the language model. This allows the application to use information that was not necessarily present in the model’s original training data. Developers should also implement access controls, source tracking, relevance evaluation, and grounding checks to improve reliability and protect internal information.

Question 344

Which Azure AI Language capability connects recognized entities to entries in a knowledge base?

1. Sentiment analysis
2. Entity linking
3. Key phrase extraction
4. Language detection

Correct Answer: 2

Explanation

Entity linking identifies recognized entities and associates them with corresponding entries in a knowledge base or reference system. This can help applications distinguish between entities that may have similar names and enrich extracted information with additional context. For example, a recognized organization or location can potentially be linked to a specific knowledge-base entry. Entity linking differs from basic named entity recognition because NER identifies entity mentions and categories, while entity linking attempts to resolve those mentions to known entities. Developers should evaluate ambiguous names and domain-specific terminology because linking accuracy depends on the available information and context.

Question 345

An AI Search application needs users to receive suggestions while typing a product name. Which feature is designed for this purpose?

1. Suggester
2. Skillset
3. Semantic caption
4. Vector field

Correct Answer: 1

Explanation

A suggester in Azure AI Search supports autocomplete and search suggestions as users enter partial text. It can improve the interactive experience by helping users discover valid terms or complete searches more quickly. Developers configure the appropriate searchable fields for suggestion behavior and should test the feature with realistic user input. Suggestions are different from semantic ranking because their primary purpose is helping users construct a query rather than ranking completed search results. They are also distinct from vector search, which retrieves semantically similar content. Proper field configuration and representative testing are important to ensure suggestions are useful and relevant.

Question 346

Which Azure AI Vision capability can identify several objects and their locations within the same image?

1. Image captioning
2. Object detection
3. OCR
4. Language detection

Correct Answer: 2

Explanation

Object detection identifies objects within an image and provides information about their locations, commonly represented using bounding regions. This makes it useful for scenarios where an application needs to know both what objects are present and approximately where they appear. Object detection differs from image classification, which assigns labels to an image or category without necessarily identifying individual object locations. OCR is focused on extracting visible text, while image captioning generates a broader textual description. Developers should evaluate object detection with images representative of the production environment because lighting, object size, occlusion, and image quality can affect results.

Question 347

A developer wants to reduce repeated model calls when an AI application receives identical frequently used requests. Which approach can improve efficiency?

1. Disable monitoring
2. Cache suitable responses
3. Increase prompt size
4. Remove authentication

Correct Answer: 2

Explanation

Caching suitable responses can reduce unnecessary model calls when an application repeatedly receives identical or sufficiently equivalent requests. This can lower latency and model usage costs while improving responsiveness. However, caching should be applied carefully when responses depend on changing information, user identity, permissions, or sensitive data. Developers should define an appropriate cache key and expiration policy and ensure that one user’s private response cannot be returned to another user. Dynamic applications may require short cache lifetimes or no caching for time-sensitive information. Performance improvements should be measured alongside correctness and freshness requirements rather than evaluated only by reduced model usage.

Question 348

Which Document Intelligence capability determines the type of a document before selecting an extraction workflow?

1. Document classification
2. Speech recognition
3. Sentiment analysis
4. Vector ranking

Correct Answer: 1

Explanation

Document classification determines which category or type a document belongs to before downstream processing occurs. This can be useful when an application receives different document types, such as invoices, contracts, purchase orders, or applications, and each type requires a different extraction model or workflow. A classifier can route documents to the appropriate processing path after identifying their category. Developers should provide representative training examples when using a custom classifier and evaluate its performance across different layouts and document variations. Classification is different from extraction because classification identifies the document type, while extraction focuses on obtaining specific fields or values.

Question 349

Which Azure AI Search query option is useful when an application must restrict results to documents matching a specific metadata condition?

1. Filter
2. Caption
3. Suggester
4. Embedding model

Correct Answer: 1

Explanation

A filter restricts search results according to structured field values stored in an Azure AI Search index. For example, an application can filter documents where Department equals Finance or where a document date falls within a specified range. Filters can be combined with keyword, semantic, hybrid, or vector retrieval. Developers should ensure that fields used in filter expressions are configured as filterable and that metadata values are consistently indexed. Filters are particularly useful for narrowing large result sets or enforcing application-level query constraints. However, authorization requirements should be enforced separately rather than relying solely on user-controlled filter expressions.

Question 350

A developer needs a model to return a fixed set of categories such as Billing, TechnicalSupport, and Sales for incoming messages. Which Azure AI Language capability is appropriate?

1. Summarization
2. Custom text classification
3. Entity linking
4. Language detection

Correct Answer: 2

Explanation

Custom text classification is designed for assigning text to categories defined by the application. In this scenario, messages can be classified into categories such as Billing, TechnicalSupport, or Sales. Developers provide labeled examples for each category so the model can learn the patterns associated with the organization’s classification requirements. The training data should contain varied examples rather than repeated versions of the same wording. After training, the model should be evaluated using separate test examples that represent real customer messages. Custom classification can then support automated routing, prioritization, workflow selection, and other downstream business processes.

Question 351

Which technique can improve the relevance of retrieved chunks by considering both the query and the complete content of candidate results?

1. Semantic reranking
2. OCR
3. Text-to-speech
4. Language detection

Correct Answer: 1

Explanation

Semantic reranking evaluates retrieved candidates using semantic relationships between the query and the candidate content. It can improve result ordering after an initial retrieval stage has produced a broader set of possible matches. This approach is useful when keyword or vector retrieval alone does not consistently place the most relevant evidence near the top. Developers should use representative evaluation queries to determine whether reranking improves retrieval quality. The initial retrieval stage should still return a sufficiently broad candidate set because reranking cannot recover documents that were never retrieved. Retrieval, ranking, filtering, and chunking should therefore be designed as a complete pipeline.

Question 352

An application receives a photograph containing a street sign and needs to extract the words printed on it. Which capability should process the image?

1. Object detection
2. OCR
3. Sentiment analysis
4. Text classification

Correct Answer: 2

Explanation

OCR is used to extract visible text from images, including signs, labels, forms, screenshots, and scanned documents. In this scenario, the street sign contains textual information that must be converted into machine-readable characters. After OCR extraction, the application can perform additional processing such as translation, entity recognition, search, or classification. OCR performance depends on factors such as image resolution, text size, orientation, contrast, and environmental conditions. Developers should test representative images and account for recognition errors when extracted text is used in important workflows. Object detection would instead focus on identifying objects and their locations.

Question 353

Which practice helps protect an AI application when retrieved documents contain instructions intended to manipulate the model?

1. Trust every retrieved instruction
2. Apply prompt-injection defenses
3. Remove all authentication
4. Increase response length

Correct Answer: 2

Explanation

Prompt-injection defenses help reduce the risk that untrusted retrieved content will manipulate a language model’s behavior. Documents used in RAG systems should be treated as data rather than automatically trusted instructions. Developers can use clear system instructions, Prompt Shields where supported, constrained tool permissions, input validation, output checks, and careful separation between instructions and retrieved content. Applications should also test adversarial documents that attempt to override system rules or cause unauthorized tool use. Security controls should be applied throughout the architecture because no single prompt instruction can guarantee protection against every form of malicious or misleading content.

Question 354

Which Azure AI Speech capability converts spoken audio into text for downstream processing?

1. Text-to-speech
2. Speech-to-text
3. SSML
4. Voice gallery

Correct Answer: 2

Explanation

Speech-to-text converts spoken audio into written text and is commonly used in transcription, voice assistants, meeting applications, and conversational systems. The resulting transcript can then be processed by other services for classification, sentiment analysis, summarization, entity extraction, or generative AI workflows. Developers should consider language selection, audio quality, microphone conditions, speaker characteristics, and domain terminology when evaluating recognition performance. Speech-to-text is fundamentally different from text-to-speech, which produces spoken audio from written text. In a conversational application, speech-to-text can serve as the input stage before language understanding or generation takes place.

Question 355

A developer wants to store document metadata such as department and publication date alongside embeddings for filtered retrieval. What index design is appropriate?

1. Vector field only
2. Metadata fields with appropriate filtering
3. Audio field only
4. Image caption only

Correct Answer: 2

Explanation

A RAG-oriented Azure AI Search index can store vector embeddings together with structured metadata fields such as department, publication date, document identifier, and access information. Appropriate metadata configuration allows the application to combine vector retrieval with filters. For example, a query can retrieve semantically similar documents while restricting results to a particular department or date range. Developers should select field types and capabilities according to how each field will be used. Metadata also supports source attribution and document management. Security-sensitive applications should additionally enforce authorization so metadata filters cannot become a mechanism for bypassing access controls.

Question 356

Which Microsoft Foundry evaluation dimension focuses on whether an AI response is supported by the supplied context?

1. Fluency
2. Groundedness
3. Latency
4. Token count

Correct Answer: 2

Explanation

Groundedness evaluates whether a generated response is supported by the information provided as context to the model. This is especially important in RAG applications because the system retrieves external evidence specifically to support generated answers. A response can be fluent and grammatically correct while still containing claims that are not supported by the retrieved information. Developers should therefore evaluate groundedness separately from other dimensions such as relevance, coherence, and safety. Representative evaluation datasets should include questions requiring different types of evidence. Monitoring groundedness after model, prompt, or retrieval changes can help identify regressions in factual support.

Question 357

Which Azure AI Search component periodically reads configured data sources and populates an index?

1. Indexer
2. Vector field
3. Suggester
4. Semantic caption

Correct Answer: 1

Explanation

An indexer automates the process of reading content from a configured data source and populating an Azure AI Search index. It can also invoke a skillset and apply field mappings during the indexing process. This is useful for applications where source data changes over time and the search index needs to be updated without manually processing every document. Developers should configure appropriate schedules, authentication, field mappings, and enrichment steps. Indexer execution should be monitored for warnings and failures because a successful run does not necessarily mean every source item was processed exactly as expected.

Question 358

A language application must summarize lengthy customer conversations into shorter text. Which Azure AI Language capability is appropriate?

1. Summarization
2. OCR
3. Object detection
4. Vector indexing

Correct Answer: 1

Explanation

Summarization reduces longer text into a shorter representation that preserves important information. It can be useful for customer conversations, documents, support cases, and meeting transcripts where users need the main points without reviewing the complete content. Developers should define whether the application requires extractive or abstractive behavior according to the supported capability and intended use. Summarization quality should be evaluated using representative conversations because important details can sometimes be omitted or expressed differently. For high-impact workflows, generated summaries should not automatically replace the original source because users may need to verify important information.

Question 359

A developer needs an AI model to follow application-wide rules such as refusing requests outside its permitted role. Where should these instructions primarily be defined?

1. User-generated content
2. System instructions
3. Search results
4. Image metadata

Correct Answer: 2

Explanation

System instructions define high-level behavioral rules and application context that should guide the model across interactions. They can establish the assistant’s role, permitted behavior, response requirements, and other application-wide constraints. User messages should not be treated as equivalent to system-level instructions because users may provide conflicting or malicious directions. Retrieved documents should also generally be treated as data rather than trusted control instructions. Developers should still use additional safeguards such as authorization, tool restrictions, content safety, and evaluation because system instructions alone cannot guarantee secure behavior. Testing should include attempts to override or bypass the intended application rules.

Question 360

An organization changes its embedding model but keeps an existing vector index configured for the old embedding dimensions. What should the team verify first?

1. Voice pronunciation
2. Vector dimension compatibility
3. Image brightness
4. Document language

Correct Answer: 2

Explanation

Vector dimensions must be compatible between the embeddings generated by the new model and the vector field configured in the search index. Changing an embedding model can therefore require changes to the index schema or creation of a new vector field or index, depending on the architecture. Existing vectors generated with the old model may also need to be regenerated so that all vectors belong to the same embedding space. Developers should verify the model’s output dimensions, index configuration, and vector-search settings before reindexing. Compatibility testing should be performed before deploying the change because mismatched dimensions can cause indexing or query failures.