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

 

Question 261

Which Microsoft Foundry capability can be used to compare an AI application’s responses against predefined evaluation criteria?

1. Evaluation
2. OCR
3. Image analysis
4. Speech synthesis

Correct Answer: 1

Explanation

Evaluation capabilities in Microsoft Foundry help developers assess AI applications against defined quality and safety criteria. Evaluations can use representative test data and metrics appropriate to the application, such as relevance, groundedness, coherence, or safety. This makes it possible to compare different prompts, models, or workflow versions using consistent inputs rather than relying only on informal manual testing. Evaluation should be performed throughout development and after significant changes because model behavior can vary when prompts, retrieved information, tools, or model versions change. Results can help identify weaknesses that require additional testing or application refinement.

Question 262

A developer needs to identify where a person, vehicle, or other object appears within an image. Which Vision capability should be selected?

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

Correct Answer: 2

Explanation

Object detection identifies objects within an image and provides their locations, commonly through bounding regions. This makes it appropriate when an application needs to determine both the type of object and where it appears. For example, a retail application could detect products on a shelf, while a safety application could identify people or vehicles in camera images. Image classification instead assigns categories to an image, while OCR focuses on extracting visible text. Developers should test object detection with images that represent actual lighting, object sizes, camera angles, backgrounds, and environmental conditions expected in the production scenario.

Question 263

Which Azure AI Search feature allows a user to restrict results to documents where a field has a particular value?

1. Filtering
2. Semantic captions
3. Vectorization
4. Index projection

Correct Answer: 1

Explanation

Filtering in Azure AI Search restricts search results according to values stored in filterable fields. For example, an application can return only documents belonging to a particular department, language, region, document type, or status. Filters can be combined with keyword, vector, hybrid, or semantic search to narrow results while retaining the chosen retrieval method. Developers must configure the relevant index field as filterable. Filtering is also useful for metadata-based retrieval in RAG applications. However, filters should be designed carefully because overly restrictive conditions can prevent relevant documents from being returned even when they otherwise match the query.

Question 264

An application receives scanned documents containing printed text. Which processing step can make that text available for search and downstream language analysis?

1. OCR
2. Face recognition
3. Speech translation
4. Object detection

Correct Answer: 1

Explanation

Optical character recognition converts visible text in images or scanned documents into machine-readable text. This extracted text can then be indexed, searched, translated, classified, summarized, or analyzed by other AI capabilities. OCR is particularly useful in document-processing and knowledge-mining pipelines where the original information is stored visually rather than as selectable text. Recognition quality depends on factors such as image resolution, orientation, layout, font, handwriting support, and background quality. Developers should evaluate OCR output before using it for critical downstream operations because recognition errors can propagate into search results or automated decisions.

Question 265

Which Azure AI Language capability is most appropriate for determining whether a customer message is positive, negative, or neutral?

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

Correct Answer: 2

Explanation

Sentiment analysis determines the emotional polarity expressed in text, commonly identifying positive, negative, or neutral sentiment and, where supported, providing more detailed sentiment information. It can help organizations analyze customer feedback, support interactions, reviews, surveys, and social content. Sentiment analysis does not primarily identify named people, places, or organizations, nor does it determine the language itself. Developers should test the service with realistic customer language because sarcasm, mixed opinions, domain-specific terminology, and short messages can affect results. Sentiment output should also be interpreted according to the application’s purpose rather than treated as an infallible human judgment.

Question 266

A RAG application needs to retrieve relevant information even when the user’s wording differs significantly from the wording in stored documents. Which technique is useful?

1. Vector search
2. Exact filtering only
3. OCR
4. Faceting

Correct Answer: 1

Explanation

Vector search can retrieve semantically similar content even when the query and stored document use different words. An embedding model converts the query and indexed content into numerical vectors, and the search system compares their semantic similarity. This is useful for natural-language questions and knowledge bases where users may describe a concept differently from the source documents. Vector search is often combined with keyword search in hybrid architectures because exact terms can remain important for identifiers and specialized terminology. Developers should evaluate embedding quality, vector dimensions, indexing configuration, and representative queries to determine whether retrieval meets application requirements.

Question 267

Which component of a generative AI prompt normally contains rules that guide how the model should behave?

1. System instructions
2. Image metadata
3. Search facet
4. Audio waveform

Correct Answer: 1

Explanation

System instructions provide high-level guidance about how a generative AI model should behave during an interaction. They can establish the assistant’s role, response style, constraints, task requirements, or other behavioral rules. Developers should write instructions clearly and avoid unnecessary ambiguity or conflicting requirements. System instructions are not a replacement for security controls because a model can still encounter malicious user input or untrusted retrieved content. Applications should therefore combine prompt design with input handling, access controls, tool restrictions, content safety measures, and evaluation. Testing should include both normal requests and attempts to override the intended behavior.

Question 268

An organization wants to classify documents into categories such as Legal, Finance, and Human Resources using its own examples. Which service capability is appropriate?

1. Custom text classification
2. OCR
3. Text-to-speech
4. Image captioning

Correct Answer: 1

Explanation

Custom text classification allows organizations to define categories and train a model using labeled examples that represent those categories. For a document-routing system, examples could be labeled Legal, Finance, Human Resources, Procurement, or other business-specific classes. The model can then classify new documents according to learned patterns. Training examples should represent the terminology, writing styles, document lengths, and variations found in production. Developers should maintain separate evaluation data to measure performance on unseen documents. Clear category definitions are also important because overlapping or inconsistent labels can make classification less reliable.

Question 269

Which Azure AI Search feature can improve retrieval for a natural-language query by using semantic understanding after initial candidate retrieval?

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

Correct Answer: 1

Explanation

The semantic ranker operates on retrieved candidates and uses language understanding to improve the ordering of results for supported semantic search scenarios. Initial retrieval can produce several documents that appear relevant based on lexical or vector similarity, but their order may not perfectly reflect the user’s intent. Semantic ranking can provide a more meaningful ordering of those candidates. It should be viewed as a ranking stage rather than the component that stores documents or performs the initial indexing operation. Developers should evaluate semantic search with realistic natural-language questions and verify that the selected semantic configuration exposes useful document content.

Question 270

An AI agent must access an enterprise API, but the application should not expose the API credential to the language model. What design is appropriate?

1. Controlled tool integration
2. Put the secret in the prompt
3. Include the key in user input
4. Publish the credential

Correct Answer: 1

Explanation

A controlled tool integration allows the agent to invoke an approved operation while keeping authentication details outside the model’s conversational context. The application or tool implementation can manage credentials securely and perform the API request on behalf of the agent. This design reduces the risk of exposing secrets through prompts, generated responses, logs, or user-visible content. Tool permissions should follow least-privilege principles, and inputs should be validated before an operation is executed. Developers should also define which actions require confirmation or additional authorization, particularly when a tool can modify records, make transactions, or access sensitive enterprise information.

Question 271

Which Document Intelligence model is designed to extract common fields from invoices?

1. Prebuilt invoice model
2. Custom NER model
3. Sentiment model
4. Vision classification model

Correct Answer: 1

Explanation

The prebuilt invoice model in Azure AI Document Intelligence is designed to extract commonly required information from supported invoice documents. Using a prebuilt model can reduce development effort when the fields required by an application are already supported. Typical invoice-processing workflows can use extracted values for validation, accounting processes, search, or downstream automation. Developers should verify the supported document types and fields against actual business requirements. If the organization needs specialized fields or document structures beyond the capabilities of the prebuilt model, a custom extraction approach may be more appropriate.

Question 272

A developer wants to use both exact keyword matching and semantic similarity for an enterprise knowledge application. Which approach should be considered?

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

Correct Answer: 2

Explanation

Hybrid search combines keyword-based retrieval with vector-based semantic retrieval. Keyword search can be particularly effective for exact names, identifiers, product codes, and specialized terminology, while vector search can find conceptually similar content despite differences in wording. Combining the approaches can therefore improve retrieval coverage for diverse enterprise queries. Azure AI Search supports hybrid scenarios where these retrieval methods contribute to the candidate result set. Developers should test different query types and evaluate relevance using realistic examples. The appropriate configuration depends on the application’s data, query patterns, embedding model, ranking requirements, and expected response latency.

Question 273

Which Azure AI Speech capability converts a recorded spoken sentence into written text?

1. Speech-to-text
2. Text-to-speech
3. Pronunciation assessment
4. Speech synthesis markup

Correct Answer: 1

Explanation

Speech-to-text converts spoken audio into written text. It is commonly used for transcription, voice commands, meeting applications, call-center processing, accessibility features, and conversational interfaces. The service processes audio and attempts to recognize the words spoken by the user. Developers should evaluate recognition using realistic speakers, accents, microphones, background noise, vocabulary, and speaking rates. Specialized application requirements may also require additional configuration or custom speech capabilities. Speech-to-text differs from text-to-speech, which generates audio from written text, while pronunciation assessment focuses on evaluating spoken pronunciation against expected reference material.

Question 274

A search index contains vectors generated by an embedding model. What must be consistent between the generated embeddings and the vector field?

1. Vector dimensions
2. File extension
3. Audio volume
4. Image resolution

Correct Answer: 1

Explanation

The dimensionality of generated embeddings must match the dimensionality configured for the corresponding vector field in the search index. If an embedding model produces vectors with a different number of dimensions than the field expects, indexing or querying cannot operate correctly. Developers should select the embedding model and index schema together and verify the required vector dimensions before creating the production index. Other vector-search settings, such as the selected algorithm and profile, must also be configured appropriately. Testing the complete indexing and query workflow can identify configuration mismatches before the application is deployed.

Question 275

Which Microsoft Foundry feature can help an AI developer inspect the execution of a multi-step workflow?

1. Tracing
2. Faceting
3. OCR
4. Image classification

Correct Answer: 1

Explanation

Tracing provides visibility into the execution of AI workflows and can help developers understand what occurred during individual steps. In a multi-step generative AI application, a trace can help reveal model calls, prompts, tool invocations, retrieved information, intermediate operations, and outputs, depending on the implementation. This information is valuable when investigating unexpected results or performance problems. Tracing should be designed with privacy and security in mind because diagnostic data may contain user input or sensitive retrieved content. Developers can use trace information alongside evaluation results and application logs to identify specific stages that require improvement.

Question 276

A company needs to identify specialized equipment names in maintenance reports. Standard entity categories do not include this business-specific entity. Which option is suitable?

1. Custom named entity recognition
2. Sentiment analysis
3. Language detection
4. Text summarization

Correct Answer: 1

Explanation

Custom named entity recognition allows an organization to define and identify entity types specific to its domain. Maintenance reports may contain specialized equipment names, internal asset identifiers, component names, or other terminology that standard entity recognition does not adequately capture. Developers train the custom model using labeled examples that identify the target entities in context. Training data should include realistic variations, abbreviations, and terminology used by employees. Evaluation should be performed on unseen reports to determine whether the model generalizes well. Custom NER can then support search, analytics, routing, extraction, and other downstream enterprise workflows.

Question 277

Which Azure AI Search property should be enabled when an application needs to use a field for user-selected category filters?

1. Filterable
2. Searchable only
3. Retrievable only
4. Hidden

Correct Answer: 1

Explanation

A field must be configured as filterable when applications need to use its values to restrict search results. Common examples include category, department, language, region, status, or document type. Filterable fields allow queries to specify conditions based on structured values. Searchable and filterable are different capabilities: a field configured for full-text searching is not automatically equivalent to a field configured for filtering. Developers should design the index schema according to how each field will be used. Appropriate field configuration is especially important in enterprise search and RAG systems that rely on metadata to narrow retrieval.

Question 278

A generative AI application must refuse or safely handle requests involving prohibited content. Which capability is relevant?

1. Content safety
2. Vector indexing
3. Semantic captions
4. OCR

Correct Answer: 1

Explanation

Content safety capabilities help applications detect and manage categories of potentially harmful content in user inputs and generated outputs. This is important for generative AI applications because both prompts and responses can contain material that violates an organization’s safety requirements. Developers should define policies appropriate to the application’s users and use cases and test the system with representative normal and adversarial inputs. Content safety should be combined with other controls, including prompt design, access management, tool restrictions, monitoring, and human review where appropriate. Safety configurations should also be reassessed when models or application behavior change.

Question 279

An organization wants to create a searchable knowledge base from documents stored in a supported data source. Which Azure AI Search component can automatically retrieve and index the source content?

1. Indexer
2. Semantic ranker
3. Vector query
4. Scoring profile

Correct Answer: 3

Explanation

An indexer automates the process of retrieving content from a supported data source and loading it into an Azure AI Search index. Depending on the configured workflow, the indexer can also invoke a skillset to enrich documents before their information is stored in the index. This makes indexers useful for recurring ingestion pipelines where source data changes over time. Developers should configure the data source, target index, field mappings, and optional enrichment pipeline correctly. Scheduling and incremental processing can also be considered when the source contains frequently changing documents that must remain available for current search experiences.

Question 280

Which practice is most useful when assessing whether an AI application performs reliably across different user scenarios?

1. Test representative scenarios
2. Test one successful prompt
3. Remove evaluation data
4. Ignore edge cases

Correct Answer: 1

Explanation

Testing representative scenarios provides a more meaningful assessment of AI application behavior than checking only a small number of successful prompts. A useful evaluation set should reflect common requests, difficult cases, edge conditions, different user wording, relevant business workflows, and important failure scenarios. Developers can then measure criteria such as relevance, correctness, groundedness, safety, latency, or task success. Representative evaluation also makes it easier to compare model or prompt changes consistently. The dataset should be maintained as application requirements evolve so that testing continues to reflect actual production conditions and newly discovered risks.