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

 

Question 361

Which Microsoft Foundry capability is used to select and deploy an available generative AI model for an application?

1. Model catalog
2. OCR skill
3. Search suggester
4. Speech recognizer

Correct Answer: 1

Explanation

The Microsoft Foundry model catalog provides access to available models that developers can evaluate and use for AI applications. It helps teams discover models according to capabilities, supported modalities, performance characteristics, and deployment considerations. After selecting an appropriate model, developers can configure a deployment and integrate it into their application. Model selection should be based on the actual workload rather than simply choosing the largest available model. Teams should consider quality, latency, cost, supported inputs and outputs, and responsible AI requirements. Evaluation against representative application data should be performed before production deployment.

Question 362

A developer needs to store the title, author, and publication date of each document in an AI Search index. Which fields should be used?

1. Vector-only fields
2. Structured index fields
3. Speech fields
4. SSML elements

Correct Answer: 2

Explanation

Structured index fields are appropriate for storing metadata such as document titles, authors, publication dates, categories, and identifiers. These fields can be configured according to how the application needs to search, filter, sort, or display the metadata. For example, a publication date may be configured for filtering and sorting, while an author field may be searchable or filterable. Metadata fields can exist alongside vector fields that store embeddings. Developers should design the index schema before loading data and verify that each field has the correct data type and capabilities for the queries the application will perform.

Question 363

Which Azure AI Search capability can improve search relevance by understanding the meaning of a query instead of relying only on lexical matches?

1. Semantic ranking
2. Blob storage
3. Field mapping
4. OCR

Correct Answer: 1

Explanation

Semantic ranking improves search relevance by using language understanding to evaluate the relationship between a query and retrieved documents. Unlike purely lexical matching, semantic ranking can recognize that different words or phrases may express related concepts. It can be particularly useful for natural-language questions and enterprise knowledge searches. Developers should configure the appropriate semantic settings and fields in the search index and evaluate the resulting ranking with representative queries. Semantic ranking can also work alongside other retrieval approaches, including keyword and vector search. The application should measure actual relevance rather than assuming semantic processing automatically improves every query.

Question 364

An AI application must determine whether a user’s request contains a person, organization, or location. Which service capability should be used?

1. Summarization
2. Named entity recognition
3. Text-to-speech
4. Image classification

Correct Answer: 2

Explanation

Named entity recognition identifies recognized entities in text and assigns them supported categories. For example, a customer message might contain a person’s name, an organization, or a geographic location. Extracted entities can support search enrichment, routing, analytics, document organization, and other downstream workflows. Developers should test NER with realistic text because spelling variations, abbreviations, uncommon names, and specialized terminology can affect results. If the application requires organization-specific entity categories, custom named entity recognition may be more suitable. Entity recognition should also be distinguished from entity linking, which attempts to associate recognized entities with entries in a knowledge base.

Question 365

Which technique can reduce the amount of retrieved text supplied to a generative model while preserving the most useful evidence?

1. Context selection
2. Voice cloning
3. Image resizing
4. Language detection

Correct Answer: 1

Explanation

Context selection reduces the amount of retrieved information supplied to a generative model by retaining the content most relevant to the user’s request. This can improve efficiency and reduce the likelihood that irrelevant information will influence the response. A RAG application can use relevance scores, semantic reranking, metadata filters, or other selection methods to determine which retrieved chunks should become model context. Developers should avoid removing evidence simply to reduce token usage because excessive compression can make answers incomplete. The selected context should be evaluated for relevance, completeness, and groundedness using representative questions and documents.

Question 366

Which Azure AI Search field capability is necessary when an application needs to sort results by publication date?

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

Correct Answer: 2

Explanation

A sortable field allows Azure AI Search to order search results according to values stored in that field. A publication date is a common example because applications may need to show the newest or oldest documents first. Developers should configure the field as sortable when creating the index schema because field capabilities determine which query operations are available. Sorting is different from filtering, which removes documents that do not satisfy a condition, and from relevance ranking, which determines how closely results match the query. Applications should select sorting behavior carefully because it can change the order users see even when relevance remains unchanged.

Question 367

A company wants to automatically identify whether uploaded documents are invoices, receipts, or purchase orders before processing them. Which capability is appropriate?

1. Document classification
2. Speech translation
3. Sentiment analysis
4. Image captioning

Correct Answer: 1

Explanation

Document classification can identify the category of an incoming document before the application selects a downstream processing workflow. In this scenario, invoices, receipts, and purchase orders can be treated as separate document classes and routed to appropriate extraction models or business processes. This is useful when a single ingestion pipeline receives multiple document types. Developers should provide representative examples for each class and evaluate the classifier against documents with different layouts and formatting. Classification should not be confused with extraction: classification determines what type of document was received, while extraction identifies specific fields and values contained within that document.

Question 368

An AI agent should access a customer’s records using the application’s identity instead of storing a password in its prompt. Which approach should be used?

1. Plain-text credentials
2. Managed identity
3. Static prompt instructions
4. Public endpoint

Correct Answer: 2

Explanation

Managed identity allows an Azure-hosted application or service to authenticate to supported resources without embedding passwords or secrets directly in prompts, source code, or configuration where they can be exposed. The identity can be granted only the permissions required to access the customer’s records or another protected resource. This supports the principle of least privilege and reduces credential-management risks. Developers should still configure authorization correctly because authentication only establishes identity. Access permissions, resource roles, logging, and data protection must also be considered. Managed identity is particularly useful when agents or application services need controlled access to Azure resources.

Question 369

Which Azure AI Vision feature can identify a rectangular region containing a detected object?

1. Object detection
2. Sentiment analysis
3. Key phrase extraction
4. Speech synthesis

Correct Answer: 1

Explanation

Object detection identifies objects in an image and provides their approximate locations, commonly through bounding regions. This allows an application to determine not only that an object exists but also where it appears within the image. Such information can support inventory systems, visual inspection, accessibility tools, and other computer vision scenarios. Object detection differs from image classification because classification generally assigns categories without identifying each object’s location. Developers should test images containing overlapping, partially hidden, or unusually sized objects because detection performance can vary according to image conditions and the characteristics of the objects being recognized.

Question 370

Which Azure AI Language capability produces a shorter representation of a lengthy document?

1. Entity linking
2. Summarization
3. Language detection
4. Custom NER

Correct Answer: 2

Explanation

Summarization produces a shorter representation of longer text while attempting to retain its most important information. It can be useful for reports, customer conversations, support cases, articles, and other documents where users need a quick overview. Developers should select the appropriate summarization approach supported by the service and evaluate results using representative content. Important details can sometimes be omitted or rephrased, so summaries should not automatically replace source documents in high-impact workflows. When summaries are used for downstream AI processing, teams should also evaluate whether the shortened content preserves the facts and context required by the next stage.

Question 371

A RAG system needs to associate every retrieved chunk with its original document and page number. What should be stored with each chunk?

1. Metadata
2. Audio samples
3. Voice settings
4. SSML markup

Correct Answer: 1

Explanation

Metadata can associate each chunk with useful source information such as the original document identifier, page number, section, URL, department, or publication date. Maintaining this information improves traceability and allows an application to provide citations or source references with generated answers. Metadata can also support filtering and document-level access controls when appropriately designed. Developers should preserve reliable identifiers during document processing and chunking so that retrieved content can be traced back to its source. Source metadata should not be treated as proof of correctness by itself; the underlying document content still needs to support the generated response.

Question 372

Which search approach is most appropriate when an application needs both exact product-code matching and semantic similarity?

1. Keyword search only
2. Vector search only
3. Hybrid search
4. Faceting only

Correct Answer: 3

Explanation

Hybrid search combines keyword retrieval with vector similarity, making it useful when an application needs both exact matching and semantic understanding. Product codes, model numbers, and identifiers often benefit from lexical matching because exact terms matter. At the same time, natural-language descriptions and conceptual queries can benefit from vector retrieval. Combining these signals gives the application multiple ways to identify relevant documents. Developers should evaluate hybrid configurations with representative queries containing both exact identifiers and natural-language requests. Ranking and filtering settings may also need adjustment because the relative contribution of keyword and vector results can influence the final ordering.

Question 373

Which Azure AI Speech capability allows an application to recognize speech from a custom vocabulary or domain?

1. Custom speech
2. Image analysis
3. Semantic ranking
4. Document classification

Correct Answer: 1

Explanation

Custom speech allows developers to adapt speech recognition for specialized terminology and domain-specific language. This can be valuable in industries where users frequently mention uncommon names, technical terms, product identifiers, or specialized vocabulary that may not be recognized reliably by a general speech model. Developers should provide representative audio and text data according to the supported customization process and evaluate the resulting model using realistic speech. Background noise, accents, speaking styles, and microphone quality should also be considered. Custom speech should be measured against a baseline so teams can determine whether customization provides meaningful improvements for their specific workload.

Question 374

A developer wants to evaluate whether a model follows instructions and completes a requested task correctly. Which evaluation dimension is most relevant?

1. Task success
2. Image resolution
3. Audio bitrate
4. Search facet count

Correct Answer: 1

Explanation

Task success evaluates whether an AI application actually completes the task specified by the user or application workflow. For example, an extraction system should return the requested fields, while an agent should perform an authorized operation according to the required instructions. This differs from fluency, which focuses on how naturally the response is written, or groundedness, which focuses on support from provided context. Developers should define task-success criteria before evaluation and use representative test cases. Edge cases and failure scenarios should also be included so that a model is not considered successful based only on simple examples.

Question 375

Which Azure AI Search configuration identifies the fields used to provide semantic search content and titles?

1. Semantic configuration
2. Managed identity
3. Storage account
4. Speech resource

Correct Answer: 1

Explanation

A semantic configuration identifies important fields that Azure AI Search can use for semantic processing, such as title, content, and keyword-related information. Selecting appropriate fields gives the semantic ranking system useful information for understanding the relationship between queries and indexed documents. Developers should ensure that these fields contain meaningful text and are configured appropriately in the search index. Semantic configuration does not replace the index itself or the underlying data source. After configuring it, developers should test natural-language queries and compare results with representative search scenarios to confirm that the selected fields support the desired relevance behavior.

Question 376

An agent should not automatically execute a financial transaction without user confirmation. Which control can enforce this requirement?

1. Human approval
2. Larger embeddings
3. More search results
4. OCR preprocessing

Correct Answer: 1

Explanation

Human approval can provide an additional safeguard before an AI agent performs a high-impact operation such as a financial transaction. The agent may prepare the requested action and present the relevant details, but the application can require an authorized person to confirm the operation before execution. This reduces the consequences of incorrect model interpretation, manipulated inputs, or unexpected tool selection. Developers should combine approval workflows with authentication, authorization, parameter validation, and logging. The approval step should clearly show the action being requested so that the reviewer can make an informed decision rather than simply approving an opaque model-generated command.

Question 377

Which Azure AI Search field capability is required when users need to apply filters based on document category?

1. Searchable
2. Filterable
3. Sortable only
4. Vectorized

Correct Answer: 2

Explanation

A field configured as filterable can be used in Azure AI Search filter expressions. If an application stores a document category such as Finance, HR, or Legal, configuring that field as filterable allows queries to restrict results to a selected category. Filtering can be combined with keyword, semantic, vector, or hybrid retrieval. Developers should ensure that category values are consistently populated and normalized because inconsistent metadata can produce unexpected results. Filter configuration should also be designed with authorization requirements in mind. A filterable field provides query control but should not be considered a substitute for server-side access enforcement.

Question 378

A developer needs to evaluate whether a RAG answer directly relates to the user’s question. Which evaluation dimension should be examined?

1. Relevance
2. Audio quality
3. OCR confidence only
4. Token encoding

Correct Answer: 1

Explanation

Relevance evaluates whether the generated response appropriately addresses the user’s request. In a RAG application, a response may be factually supported by retrieved documents but still fail to answer the actual question. Developers should therefore evaluate relevance separately from groundedness. Representative evaluation datasets should contain straightforward questions, ambiguous requests, multi-part questions, and cases where the correct behavior is to acknowledge insufficient information. Measuring relevance across these scenarios helps identify prompt, retrieval, and model problems. Teams can compare relevance results before and after changes to prompts, retrieval settings, models, or context-selection strategies.

Question 379

Which Microsoft Foundry component can be used to test different prompt variations without rebuilding the entire application?

1. Prompt flow
2. Blob container
3. Speech recognizer
4. OCR engine

Correct Answer: 1

Explanation

Prompt flow supports structured development and testing of AI workflows that may contain prompts, models, tools, and processing steps. Developers can create and compare variations in prompts or workflow logic while using controlled inputs and evaluation data. This is useful when optimizing an application because prompt changes can affect relevance, groundedness, safety, cost, and response consistency. Teams can evaluate different variants before integrating a preferred configuration into a production workflow. Developers should preserve representative test cases and track changes so that improvements can be distinguished from changes caused by different input data or model versions.

Question 380

A production AI application suddenly begins producing less relevant answers after a retrieval configuration change. What should the team compare first?

1. Evaluation results before and after the change
2. Screen brightness
3. Keyboard layout
4. Audio speaker volume

Correct Answer: 1

Explanation

Comparing evaluation results before and after a retrieval configuration change provides evidence about whether the change introduced a measurable regression. A representative evaluation dataset can be used to compare dimensions such as retrieval relevance, groundedness, task success, and response quality. Developers should also inspect retrieved documents for failed examples because a retrieval change may alter which evidence reaches the model. Keeping evaluation baselines is important for detecting regressions after modifications to ranking, filtering, embeddings, chunking, or search configuration. The investigation should use actual failed cases rather than relying only on subjective impressions from a small number of responses.