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

 

Question 221

Which Azure AI Search feature allows results to be restricted based on a document’s metadata?

1. Semantic captions
2. Filtering
3. Vectorization
4. Semantic ranking

Correct Answer: 2

Explanation

Filtering in Azure AI Search allows applications to restrict search results according to values stored in structured fields. For example, a company can filter documents by department, region, document status, category, or access level. The relevant field must be configured as filterable within the index. Filters can be combined with keyword, vector, or hybrid search to narrow the set of results while preserving relevance-based retrieval. In enterprise applications, filtering can improve both usability and retrieval precision. However, filters should not replace application-level authorization because access to sensitive information must be controlled through appropriate security mechanisms.

Question 222

A developer wants to identify whether text is written in Spanish, French, or English before processing it. Which Azure AI Language capability is appropriate?

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

Correct Answer: 2

Explanation

Language detection identifies the language represented in a text input. It is useful when an application receives multilingual content without knowing the language beforehand. After determining the language, the application can route the content to an appropriate translation, classification, sentiment, or other language-processing workflow. For example, a global customer-support platform can detect a message’s language before selecting a translation path. Developers should test the capability using realistic messages because very short inputs, mixed-language text, unusual spelling, or ambiguous content can make language identification more difficult. The supported language set should also match application requirements.

Question 223

A RAG system retrieves documents from Azure AI Search before sending information to a generative model. What is the primary purpose of the retrieval step?

1. Increase image resolution
2. Provide relevant context
3. Convert speech to audio
4. Create training labels

Correct Answer: 2

Explanation

The retrieval step in a RAG system provides relevant external context to the generative model. Instead of relying entirely on information learned during model training, the application searches a trusted knowledge source and supplies useful documents or chunks with the user’s request. This can improve responses involving private, current, or organization-specific information. Retrieval quality is critical because irrelevant or incomplete context can lead to poor answers even when the model itself is capable. Developers should evaluate indexing, chunking, embeddings, filtering, ranking, and query formulation to ensure that the retrieval stage consistently provides useful evidence.

Question 224

Which Azure AI Document Intelligence capability is designed for extracting information from receipts without creating a model from scratch?

1. Custom extraction
2. Prebuilt receipt model
3. Speech recognition
4. Sentiment analysis

Correct Answer: 2

Explanation

The prebuilt receipt model in Azure AI Document Intelligence is designed to extract commonly required information from supported receipt documents. This can reduce development effort because the organization does not need to train a custom model for standard receipt fields. Depending on the supported model, extracted information can include merchant details, transaction dates, totals, and other relevant fields. A prebuilt model is appropriate when its supported schema meets the application’s requirements. If an organization needs specialized fields or must process a significantly different document structure, a custom extraction model may be more suitable.

Question 225

Which Azure AI Search capability improves the ordering of retrieved results using language understanding?

1. Semantic ranker
2. Indexer
3. Data source
4. Facet

Correct Answer: 1

Explanation

The semantic ranker improves the ordering of candidate search results by applying language understanding to the query and retrieved content. Traditional keyword or vector retrieval identifies candidate documents, but the most relevant result is not always ranked first. Semantic ranking can analyze the relationship between the query and candidate content and provide a more meaningful ordering. It can be useful for natural-language questions, enterprise search, and knowledge applications. Developers should configure semantic search appropriately and evaluate its effectiveness using representative queries. Semantic ranking is a ranking stage and does not replace the search index or underlying retrieval process.

Question 226

A voice application must translate a user’s spoken German into English text. Which Azure AI capability is most appropriate?

1. OCR
2. Speech translation
3. Text classification
4. Object detection

Correct Answer: 2

Explanation

Speech translation is designed for scenarios where spoken language must be recognized and translated into another language. Azure AI Speech can process the spoken source language and provide translated output in a selected target language. This can support multilingual customer service, meetings, travel applications, and interactive voice solutions. Speech translation differs from ordinary speech-to-text because its primary purpose includes language translation rather than simply transcribing the original speech. Developers should configure the correct source and target languages and evaluate performance using realistic audio, accents, background conditions, and speaking styles expected in the production environment.

Question 227

Which Azure AI Language capability is useful for finding important concepts in a large collection of customer comments?

1. Key phrase extraction
2. Speech synthesis
3. Image classification
4. Custom entity recognition

Correct Answer: 1

Explanation

Key phrase extraction identifies significant terms and concepts within unstructured text. It can help organizations analyze large volumes of customer comments without requiring every message to be manually reviewed. The extracted phrases can support analytics, search indexing, content organization, and subsequent processing. For example, a business might discover recurring concepts related to shipping, product quality, pricing, or support. Key phrase extraction does not assign predefined categories or determine whether a comment is positive or negative. Developers can combine it with sentiment analysis or custom classification when the application needs additional information about the text.

Question 228

A developer needs to generate an embedding for each document chunk before storing the chunks in a vector index. What is the embedding used for?

1. Semantic similarity
2. Password storage
3. Audio compression
4. Image cropping

Correct Answer: 1

Explanation

An embedding represents content as a numerical vector that captures semantic characteristics. When document chunks are converted into embeddings, those vectors can be stored in a vector-enabled search index. A user query can also be converted into an embedding, allowing the search system to compare the query with document vectors and retrieve semantically similar content. This is a core component of many RAG architectures. Embeddings do not encrypt documents or replace the original source text. Developers should ensure that the embedding model and vector field configuration are compatible, including required vector dimensions and search settings.

Question 229

An AI agent must perform an approved operation in an enterprise application when requested by a user. Which component provides the connection to that application?

1. Tool
2. Semantic caption
3. Vector field
4. OCR skill

Correct Answer: 1

Explanation

A tool provides an agent with a controlled way to interact with an external service or application. For example, an enterprise agent could use a tool to retrieve an order, create an approved support ticket, or query an internal system. The tool defines the operations available to the agent and can pass structured inputs to the connected service. Developers should restrict tool permissions, validate inputs, handle errors, and enforce authentication and authorization. This separation between the model and external actions is important because the model should not receive unrestricted access to enterprise systems simply because it can generate natural-language instructions.

Question 230

Which Azure AI Vision capability is appropriate when an application needs to extract printed text from a photograph?

1. Object detection
2. Image captioning
3. OCR
4. Sentiment analysis

Correct Answer: 3

Explanation

Optical character recognition extracts visible text from images such as photographs, scanned pages, screenshots, and other visual content. It allows applications to convert text that exists in an image into machine-readable information for subsequent processing. After OCR, the extracted text can be searched, translated, classified, summarized, or passed to a generative AI model. OCR should not be confused with object detection, which identifies objects and their locations, or image captioning, which generates descriptions of visual content. Image quality, text orientation, background complexity, and font characteristics can influence recognition accuracy.

Question 231

A company wants to create a custom classifier for support tickets using categories defined by its own business. What type of training data is required?

1. Unlabeled audio
2. Labeled text
3. Unlabeled images
4. Video frames

Correct Answer: 2

Explanation

Custom text classification requires labeled text examples that associate each training document or message with the intended business category. For example, support tickets might be labeled Billing, Account Access, Technical Issue, or Product Information. The model learns patterns from these examples and applies them when classifying new messages. The quality of the labels is important because inconsistent category definitions can make the model difficult to train effectively. Training data should also represent the language, terminology, message length, and variation expected in production. A separate evaluation dataset should be used to measure how well the model generalizes.

Question 232

Which Azure AI Search approach can retrieve documents that contain exact terms while also finding semantically similar content?

1. Hybrid search
2. OCR
3. Faceting only
4. Speech recognition

Correct Answer: 1

Explanation

Hybrid search combines keyword-based retrieval with vector similarity search. Keyword retrieval is useful when exact terms, product identifiers, names, or technical expressions matter. Vector retrieval can find documents that express similar concepts using different wording. Combining the two approaches allows an application to handle both precise and natural-language queries. Azure AI Search can also apply ranking and filtering capabilities to refine the results. Developers should test hybrid configurations using representative queries because the balance between lexical and semantic signals can affect retrieval quality. The objective is to provide relevant evidence for downstream applications such as enterprise search or RAG.

Question 233

Which security mechanism can allow an Azure-hosted application to access supported resources without storing credentials directly in source code?

1. Managed identity
2. Shared password
3. Hard-coded API key
4. Anonymous access

Correct Answer: 1

Explanation

Managed identity provides an Azure-based identity that applications can use to authenticate to supported services without embedding long-lived credentials in source code. Administrators can assign appropriate permissions to the identity through supported authorization mechanisms. This reduces the need to manage passwords or secrets within application code and can improve credential-management practices. Managed identity should still be configured with least-privilege permissions because authentication does not automatically determine what actions the application is authorized to perform. Developers should also consider network restrictions, auditing, secret rotation requirements for other credentials, and the specific authentication capabilities of the target service.

Question 234

A generative AI system sometimes answers questions using information that is not present in its retrieved documents. Which evaluation measure is especially useful?

1. Groundedness
2. Image accuracy
3. Speech volume
4. OCR resolution

Correct Answer: 1

Explanation

Groundedness evaluates whether a generated response is supported by the information provided as context. This is especially important in RAG applications because retrieved documents are intended to provide evidence for the model’s response. A model can produce fluent and convincing text while still introducing unsupported information, so groundedness evaluation helps identify this problem. Developers can create test cases where the correct answer is explicitly supported by source material and cases where the requested information is absent. Groundedness should be evaluated alongside retrieval relevance, response quality, safety, and other application-specific criteria.

Question 235

Which Azure AI Speech technology can evaluate a learner’s spoken pronunciation against reference text?

1. Speech-to-text
2. Pronunciation assessment
3. Text-to-speech
4. Language detection

Correct Answer: 2

Explanation

Pronunciation assessment evaluates spoken language against expected pronunciation and can provide feedback useful for language-learning applications. A learner can read or repeat reference text while the application captures the spoken response for analysis. The capability can assess pronunciation-related characteristics supported by the service and configuration. It is different from ordinary speech-to-text, which primarily converts speech into written words, and from text-to-speech, which generates spoken audio from text. Developers should evaluate the feature with appropriate reference text, languages, speakers, microphones, and recording environments to ensure the results meet the application’s educational requirements.

Question 236

A developer wants to compare a new AI model with an existing model using the same production-like questions. What should be created?

1. Representative evaluation dataset
2. New image index
3. Speech profile
4. Storage container only

Correct Answer: 1

Explanation

A representative evaluation dataset provides consistent test inputs for comparing different models or application configurations. The dataset should reflect realistic production scenarios, including common user requests, difficult questions, edge cases, and important business workflows. Using identical inputs allows developers to compare outputs according to predefined criteria such as relevance, correctness, groundedness, safety, or response quality. Evaluation results can then be considered alongside latency, cost, supported features, and operational requirements. A representative dataset should be maintained and updated as application requirements change so that evaluations continue to reflect the scenarios that matter to users.

Question 237

Which Azure AI Search feature can return grouped counts such as the number of documents in each category?

1. Faceting
2. Vector search
3. Semantic ranker
4. OCR

Correct Answer: 1

Explanation

Faceting allows Azure AI Search applications to display counts grouped by values in selected fields. For example, a search portal might show how many matching documents belong to each department, file type, region, or category. Users can then select a facet to narrow the search results. The field used for faceting must be configured appropriately in the index. Faceting is primarily an exploration and filtering aid rather than a relevance-ranking mechanism. It can be combined with keyword, vector, hybrid, and semantic search to provide users with a more flexible way to navigate large collections of indexed information.

Question 238

An organization needs to identify proprietary product names that are not recognized by the standard entity model. Which capability should be considered?

1. Custom named entity recognition
2. Sentiment analysis
3. Language detection
4. Key phrase extraction

Correct Answer: 1

Explanation

Custom named entity recognition is designed for organizations that need to identify specialized entity types not adequately covered by standard language models. Proprietary product names, internal project identifiers, specialized equipment names, or organization-specific terminology can require a custom approach. Developers provide labeled examples showing the target entities in context, and the resulting model can identify similar entities in new text. The training dataset should contain enough variation to represent how the terminology appears in real documents. Evaluation should use separate examples so developers can determine whether the custom model generalizes effectively beyond its training data.

Question 239

A RAG application retrieves ten chunks, but only three are strongly relevant to the user’s question. What can help before sending the context to the model?

1. Re-ranking
2. Speech synthesis
3. Image generation
4. OCR conversion

Correct Answer: 1

Explanation

Re-ranking can improve the ordering of retrieved content so that the most relevant chunks appear first. In a RAG pipeline, the initial retrieval stage may return several candidates, including some that are only weakly related to the query. A ranking stage can analyze the relationship between the query and candidate content and prioritize stronger matches. The application can then select an appropriate number of top results before constructing the model context. This can reduce unnecessary context and improve grounding. Developers should evaluate the ranking approach using representative queries and determine whether important evidence remains available after selection.

Question 240

Which principle requires an AI application to provide only the permissions necessary for its assigned task?

1. Full trust
2. Least privilege
3. Anonymous access
4. Shared administration

Correct Answer: 2

Explanation

The principle of least privilege requires an application, identity, or agent to receive only the permissions needed to perform its authorized task. For example, an AI assistant that only needs to read customer records should not receive permission to delete or modify those records. Limiting permissions reduces the potential impact of compromised credentials, incorrect tool calls, prompt manipulation, and software defects. Least privilege should be implemented through appropriate authentication and authorization mechanisms and reviewed periodically. In agent-based systems, it is especially important because tools may allow the model to perform actions against real enterprise systems rather than simply generate text.