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

 

Question 321

Which Microsoft Foundry feature allows an agent to maintain conversation context across multiple user messages?

1. Agent thread
2. Search index
3. OCR pipeline
4. Vision model

Correct Answer: 1

Explanation

An agent thread provides a mechanism for maintaining conversation context across multiple interactions with an AI agent. Instead of treating every user message as an isolated request, the application can associate related messages with the same conversational thread. This helps the agent understand references to earlier exchanges and maintain continuity. Developers should manage thread data carefully because long conversations can increase context size and cost. Applications should also avoid retaining unnecessary sensitive information. When persistent context is not required, developers can use shorter interaction scopes or selectively summarize previous messages to keep the conversation efficient.

Question 322

An application must allow users to search for products by meaning rather than exact keywords. Which capability should be implemented?

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

Correct Answer: 2

Explanation

Vector search allows an application to retrieve content according to semantic similarity rather than requiring exact keyword matches. Product searches can therefore return relevant items even when the user’s wording differs from the terminology stored in the catalog. For example, a query describing a lightweight device for travel may retrieve products whose descriptions use different wording. The application typically converts product information and user queries into embeddings before performing similarity search. Developers should select an appropriate embedding model and ensure that vector dimensions match the configured index fields. Retrieval quality should be validated with realistic product queries.

Question 323

A developer needs to enrich indexed documents with automatically extracted entities before storing additional information. Which Azure AI Search component is designed for this purpose?

1. Skillset
2. Scoring profile
3. Suggester
4. Vector profile

Correct Answer: 1

Explanation

An Azure AI Search skillset defines enrichment operations that can process content during indexing. Skills can perform tasks such as text splitting, OCR, entity recognition, language detection, or other supported enrichment operations. The resulting enriched information can then be mapped into searchable or filterable fields or used in downstream processing. Skillsets are particularly useful when source documents contain unstructured information that must be transformed before search. Developers should design the enrichment pipeline according to the required output schema and verify field mappings carefully. Monitoring indexing failures is also important because an unsuccessful enrichment step can affect the resulting searchable content.

Question 324

A team needs to control how a search query ranks documents containing specific business terms. Which Azure AI Search feature should they configure?

1. Semantic caption
2. Scoring profile
3. Speech configuration
4. Document classifier

Correct Answer: 2

Explanation

A scoring profile allows developers to influence the ranking of search results by applying custom scoring logic to indexed fields. For example, an application can give additional importance to documents containing values in a particular field or boost results according to business-specific criteria. Scoring profiles are useful when default relevance ranking does not fully represent the application’s requirements. Developers should test ranking changes with representative queries because aggressive boosting can cause less relevant documents to appear above more relevant ones. Scoring profiles are different from filters, which restrict which documents can appear rather than primarily changing their ranking.

Question 325

Which Microsoft Foundry capability can help developers compare several model responses against the same evaluation dataset?

1. Model evaluation
2. Speech translation
3. Image tagging
4. OCR

Correct Answer: 1

Explanation

Model evaluation provides a structured method for assessing AI outputs against defined datasets and evaluation criteria. Developers can submit the same representative inputs to different models or model versions and compare dimensions such as relevance, coherence, groundedness, fluency, safety, or task success. This is useful when selecting a model for a production workload because model quality should be measured against the application’s actual requirements rather than assumed from general benchmarks. Evaluation datasets should contain realistic examples and important edge cases. Developers should also track evaluation results over time so that regressions can be identified after model, prompt, or retrieval changes.

Question 326

A document-processing solution receives PDFs containing scanned pages with no selectable text. Which capability can extract the visible characters?

1. Sentiment analysis
2. Entity linking
3. OCR
4. Text classification

Correct Answer: 3

Explanation

Optical character recognition, or OCR, extracts text from images and scanned documents. It is particularly useful when a PDF contains page images rather than an embedded text layer. Once text has been extracted, the application can perform additional processing such as classification, entity recognition, search indexing, summarization, or information extraction. OCR accuracy can vary according to image quality, font characteristics, handwriting, orientation, and document layout. Developers should test representative documents and consider preprocessing when appropriate. OCR should not be confused with document classification, which determines document categories, or entity recognition, which identifies specific entities within already available text.

Question 327

An AI application needs to prevent users from accessing search results belonging to another department. What should the application combine with search filtering?

1. Authorization controls
2. Image captions
3. Speech synthesis
4. Prompt formatting

Correct Answer: 1

Explanation

Search filtering can restrict results according to metadata, but authorization controls are required to enforce who is actually allowed to access information. For example, an application may filter documents by department while also checking the authenticated user’s permissions before executing a query. Relying only on client-controlled filters can create security risks because a user might attempt to alter the filter and retrieve restricted records. Developers should enforce authorization on the server side and design indexes with appropriate security metadata. Sensitive applications should also test unauthorized access scenarios to confirm that users cannot bypass the intended document-level restrictions.

Question 328

Which Azure AI Language capability identifies important concepts or terms within a document without requiring predefined entity categories?

1. Sentiment analysis
2. Key phrase extraction
3. Custom NER
4. Language detection

Correct Answer: 2

Explanation

Key phrase extraction identifies important concepts and phrases that summarize the main topics of a text. It can help applications analyze customer feedback, organize documents, generate metadata, or identify recurring subjects. Unlike named entity recognition, key phrase extraction does not require the developer to define specific entity categories such as person, organization, or location. The results can be used as additional metadata for search and analytics. Developers should remember that key phrases are not necessarily complete summaries and should not automatically be treated as authoritative classifications. Testing with representative documents helps determine whether extracted phrases meet the application’s intended use.

Question 329

A RAG application retrieves five documents for a user query but only two contain useful evidence. What should the application do before generating the final answer?

1. Increase audio volume
2. Remove all metadata
3. Select relevant context
4. Disable embeddings

Correct Answer: 3

Explanation

The application should select relevant context from the retrieved documents before passing information to the language model. Including every retrieved result can introduce irrelevant material, consume additional context space, and potentially make the generated response less focused. A retrieval pipeline can use ranking, relevance thresholds, metadata filters, or semantic reranking to identify the strongest evidence. Developers should also preserve enough context to answer the question accurately. The objective is not simply to minimize the number of documents but to provide the model with useful, trustworthy information that directly supports the requested response.

Question 330

Which Azure AI Search feature allows users to narrow results by values such as category or department?

1. Facets
2. SSML
3. Embeddings
4. Speech recognition

Correct Answer: 1

Explanation

Facets allow applications to present grouped counts based on structured field values, helping users narrow search results interactively. For example, an online catalog can display counts for categories, brands, or departments and allow the user to select one of those values. Facets work with appropriate filterable fields and are especially useful in search interfaces where users need to refine broad result sets. Developers should select meaningful metadata fields and ensure their values are consistently indexed. Facets do not themselves determine semantic relevance; they provide a structured way to summarize and refine the available search results.

Question 331

A developer wants an AI agent to call a weather API only when current weather information is requested. What should be configured?

1. Agent tool
2. Static embedding
3. OCR skill
4. Search suggester

Correct Answer: 1

Explanation

An agent tool allows an AI agent to interact with an external capability such as a weather API. The tool can be described so that the agent knows when it is appropriate to invoke it and what parameters are required. Because weather information changes continuously, retrieving current data through an authorized API is more appropriate than relying on static model knowledge. Developers should validate tool arguments, authenticate API access, handle failures, and limit permissions. Tool descriptions should clearly explain their purpose so the agent can make appropriate decisions about when a tool should be called during an interaction.

Question 332

Which Azure AI Speech feature allows developers to control pauses, emphasis, and pronunciation in synthesized speech?

1. Vector search
2. SSML
3. OCR
4. Semantic ranking

Correct Answer: 2

Explanation

Speech Synthesis Markup Language, or SSML, provides structured controls for influencing how text is synthesized into speech. Developers can use supported SSML elements to control aspects such as pauses, emphasis, pronunciation, speaking rate, pitch, and other speech characteristics. This can improve the naturalness and clarity of generated audio, especially for applications involving announcements or specialized terminology. Developers should test SSML with the selected voice because supported features can vary. SSML affects speech synthesis and does not perform speech recognition. It should therefore be applied when converting application text into controlled spoken output.

Question 333

A developer wants an AI Search indexer to process files stored in Azure Blob Storage. Which component identifies where the source files are located?

1. Data source
2. Scoring profile
3. Semantic configuration
4. Evaluation dataset

Correct Answer: 1

Explanation

An Azure AI Search data source defines the external repository from which an indexer retrieves content. For files stored in Azure Blob Storage, the data source contains the necessary connection and source information used by the indexing process. The indexer then reads the source and can apply field mappings or a skillset before populating the search index. Developers should configure appropriate authentication and ensure that the identity used by the indexer has sufficient permissions. A data source differs from an index, which stores searchable content, and an indexer, which controls the process of reading and processing source data.

Question 334

A team wants to identify whether generated answers contain unsafe content before displaying them to users. Which capability should be included?

1. Content safety
2. Vector dimensions
3. Faceting
4. Document chunking

Correct Answer: 1

Explanation

Content safety capabilities can help identify potentially harmful or inappropriate material in AI inputs and outputs. For a generative application, output safety checks can be incorporated before responses are presented to users or passed into additional automated processes. Developers should define safety requirements according to the application’s audience and use case and evaluate the system with representative harmful and benign examples. Safety filtering should not be considered the only protection. Applications should also use appropriate access controls, prompt defenses, monitoring, and human review for high-impact scenarios. Safety behavior should be tested after model or prompt changes.

Question 335

Which approach is most appropriate for preparing long documents for a RAG system?

1. Store the entire document in every prompt
2. Split documents into meaningful chunks
3. Remove all document metadata
4. Convert every page to audio

Correct Answer: 2

Explanation

Splitting long documents into meaningful chunks helps a RAG system retrieve focused portions of content instead of supplying an entire document for every question. Effective chunking considers logical boundaries such as sections, paragraphs, or related topics while preserving enough surrounding context to maintain meaning. Each chunk can be embedded and stored with metadata such as document identity, section, or source location. Poor chunking can reduce retrieval quality by separating related information or creating excessively large passages. Developers should test different chunking strategies using representative questions and measure whether the retrieved chunks contain the evidence needed to generate accurate responses.

Question 336

A developer wants to identify references to people, organizations, and locations in customer messages. Which capability is appropriate?

1. Named entity recognition
2. Text-to-speech
3. Image captioning
4. Speech translation

Correct Answer: 1

Explanation

Named entity recognition identifies entities in text according to supported entity categories. Common examples include people, organizations, locations, dates, and other recognized entity types. This capability can support document organization, search enrichment, analytics, routing, and information extraction. Developers should distinguish standard named entity recognition from custom NER, which is used when organization-specific entity categories are required. Entity extraction should be evaluated using representative business language because unusual names, abbreviations, spelling variations, and specialized terminology can affect results. Extracted entities should also be validated before being used for sensitive automated decisions or actions.

Question 337

Which Azure AI Search configuration determines which fields participate in semantic ranking?

1. Semantic configuration
2. Storage account
3. Speech voice
4. Prompt template

Correct Answer: 1

Explanation

A semantic configuration specifies the fields that Azure AI Search should use when performing semantic processing and ranking. Developers can identify appropriate title, content, and keyword-oriented fields so the search service has useful information for interpreting the query and documents. Proper configuration can improve the relevance of semantic search results and related semantic features. Developers should ensure that selected fields contain meaningful content and that the index schema supports the intended search scenario. Semantic configuration is separate from vector configuration, although both can be used in a broader retrieval architecture when an application requires semantic and vector-based search.

Question 338

An AI solution needs to translate user messages while preserving the original meaning across supported languages. Which Azure service should be used?

1. Azure AI Translator
2. Azure AI Vision
3. Azure AI Speech only
4. Azure AI Document Intelligence

Correct Answer: 1

Explanation

Azure AI Translator provides machine translation capabilities for converting text between supported languages. It can be integrated into multilingual applications such as customer-support systems, content-processing workflows, and communication platforms. Developers should select the correct source and target languages and consider whether specialized translation requirements need additional configuration. Translation quality should be tested with domain-specific terminology because general-purpose machine translation may not always handle specialized language perfectly. Translator can also be combined with other Azure AI services when an application needs to detect language, analyze translated content, or convert translated text into speech for a multilingual user experience.

Question 339

Which evaluation practice provides the strongest indication of how an AI application will perform in its intended business scenario?

1. Testing only one example
2. Using representative test data
3. Measuring only response length
4. Checking model name

Correct Answer: 2

Explanation

Representative test data provides a stronger basis for evaluating an AI application because it reflects the types of inputs, terminology, edge cases, and user expectations that the production system will encounter. A small or unrealistic test set can make an application appear successful while missing important failure patterns. Developers should include common scenarios, difficult cases, ambiguous inputs, and safety-sensitive examples where relevant. Evaluation should use clearly defined criteria such as relevance, groundedness, correctness, safety, or task success. Results can then be compared across prompt, model, retrieval, and configuration changes to identify meaningful improvements or regressions.

Question 340

An AI agent can send emails through an external tool. Which control is most appropriate before allowing high-impact email actions?

1. Remove authentication
2. Require appropriate authorization or approval
3. Grant administrator permissions
4. Disable logging

Correct Answer: 2

Explanation

High-impact agent actions such as sending emails should be protected with appropriate authorization and, when necessary, an approval step. An AI model can select a tool based on its interpretation of user instructions, but applications should not rely solely on the model to determine whether an action is safe or authorized. Developers can restrict recipients, validate parameters, require confirmation for sensitive operations, and enforce identity-based permissions. Logging should also record important tool activity for auditing and troubleshooting. These controls reduce the consequences of incorrect model decisions, malicious inputs, prompt injection, or unauthorized attempts to use the agent.