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

 

Question 121

Which Azure AI Search feature is used to combine keyword matching with vector similarity search?

1. Faceting
2. Hybrid search
3. Synonym maps
4. Scoring profiles

Correct Answer: 2

Explanation

Hybrid search combines traditional keyword-based retrieval with vector search in Azure AI Search. Keyword search is useful when exact terms, names, or specific phrases matter, while vector search retrieves documents based on semantic similarity between embeddings. Combining both approaches can improve retrieval quality when user queries contain both precise terminology and conceptual meaning. The search results can also be processed with semantic ranking to improve their relevance. This approach is particularly useful in retrieval-augmented generation applications where the system needs to retrieve relevant enterprise information before sending context to a generative AI model.

Question 122

A developer needs an Azure AI Search vector index that performs approximate nearest neighbor searches efficiently across a large collection. Which algorithm is commonly used?

1. HNSW
2. BM25
3. TF-IDF
4. PageRank

Correct Answer: 1

Explanation

HNSW, or Hierarchical Navigable Small World, is a commonly used approximate nearest neighbor algorithm for vector search in Azure AI Search. It creates a graph-based structure that allows the service to efficiently locate vectors that are close to a query vector without comparing the query against every stored vector. This improves search performance when indexes contain large numbers of embeddings. Approximate nearest neighbor methods generally provide a strong balance between retrieval speed and relevance. The choice of vector search configuration should consider the expected dataset size, latency requirements, and desired retrieval accuracy for the application.

Question 123

An AI application must prevent users from submitting prompts containing known attack patterns designed to manipulate a generative model. Which capability should be considered?

1. Image tagging
2. Speech synthesis
3. Prompt Shields
4. Key phrase extraction

Correct Answer: 3

Explanation

Prompt Shields are designed to help detect and mitigate certain prompt injection and related attacks against generative AI applications. Attackers may attempt to manipulate model instructions by embedding malicious directions within user input or retrieved content. Protecting the application requires evaluating untrusted inputs and controlling how external content is interpreted by the model. Prompt Shields can form part of a broader defense strategy that includes input validation, access controls, grounded responses, content filtering, and careful system instructions. They should not be treated as the only security mechanism in a production generative AI solution.

Question 124

A company wants an AI Search index to allow users to filter results by department while still performing vector retrieval. What should the index contain?

1. A filterable department field
2. A speech recognition model
3. A sentiment classifier
4. An image generation model

Correct Answer: 1

Explanation

A filterable department field allows Azure AI Search to restrict results according to metadata while vector retrieval is performed. For example, an enterprise knowledge application could store a department value such as Finance, Human Resources, or Engineering alongside each document or chunk. A query can then retrieve semantically similar vectors while limiting results to the selected department. This is useful for improving relevance and enforcing logical content boundaries. Proper index schema design is important because fields intended for filtering must be configured with the appropriate searchable, filterable, sortable, or facetable attributes according to the application’s requirements.

Question 125

Which Microsoft Foundry capability helps developers test prompts and compare model responses during application development?

1. Prompt flow
2. Face detection
3. OCR
4. Speech translation

Correct Answer: 1

Explanation

Prompt flow provides capabilities for developing, testing, evaluating, and organizing workflows involving prompts and AI models. Developers can use it to connect prompts, models, tools, and other processing steps while examining outputs during development. This makes it useful when refining an AI application because different prompts, models, or workflow configurations can be tested systematically. Evaluation can also be incorporated to assess application quality against defined criteria. Instead of manually testing every interaction independently, developers can use structured workflows and evaluation approaches to identify weaknesses and improve the behavior of generative AI applications before production deployment.

Question 126

An organization wants an AI agent to access an internal inventory system through an approved API. What should be configured for the agent?

1. A tool
2. An image tagger
3. A speech recognizer
4. A translation dictionary

Correct Answer: 1

Explanation

An agent can use a tool to interact with an external system such as an inventory API. Tools extend an agent beyond generating text by allowing it to retrieve information or perform authorized actions. For example, an inventory tool could accept a product identifier and return current stock information from an enterprise application. The tool should expose only the operations required by the agent and should use appropriate authentication and authorization. This approach helps separate the model’s reasoning from the actual business operation. Tool access should also be monitored and restricted according to the principle of least privilege.

Question 127

Which Azure AI Search component can improve the ordering of search results by using language understanding?

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

Correct Answer: 2

Explanation

The semantic ranker in Azure AI Search can improve the ordering of retrieved search results by applying language understanding to the candidate documents. Traditional keyword or vector retrieval may return several relevant documents, but their order is not always optimal for the user’s question. Semantic ranking can analyze the relationship between the query and retrieved content and produce a more useful ranking. This is especially valuable in question-answering and knowledge-search applications. Semantic ranking works as a ranking stage rather than replacing the underlying index, data source, or retrieval mechanisms used to obtain candidate documents.

Question 128

A developer needs to extract structured fields from standardized invoices without training a custom model. Which service capability is appropriate?

1. Prebuilt invoice model
2. Object detection
3. Custom text classification
4. Speech synthesis

Correct Answer: 1

Explanation

The prebuilt invoice model in Azure AI Document Intelligence is designed to extract commonly required invoice information without requiring an organization to create and train a custom extraction model from scratch. Depending on the document, extracted information can include fields such as vendor details, invoice identifiers, dates, totals, and line items. Using a prebuilt model can accelerate development when the document type is already supported. A custom model may be more appropriate when an organization needs specialized fields or document structures that are not adequately handled by available prebuilt capabilities.

Question 129

A speech application must pronounce words according to a specific pronunciation and speaking style. Which technology can provide detailed control over speech synthesis?

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

Correct Answer: 1

Explanation

Speech Synthesis Markup Language, or SSML, provides developers with structured controls for speech synthesis. It can be used to influence aspects such as pronunciation, pauses, emphasis, speaking rate, pitch, and other characteristics supported by the speech service. This is useful for applications such as virtual assistants, educational systems, accessibility solutions, and interactive voice experiences where default pronunciation or delivery may not be sufficient. SSML allows the application to provide additional instructions to the speech synthesis engine while keeping the generated speech understandable and appropriately formatted for the intended user experience.

Question 130

A team wants to identify customer-specific names of products and internal systems that are not reliably recognized by a pretrained language model. Which Azure AI Language capability can address this requirement?

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

Correct Answer: 1

Explanation

Custom named entity recognition allows an organization to identify domain-specific entities that may not be consistently recognized by a general pretrained model. For example, a company might need to identify internal product codes, proprietary system names, specialized equipment, or industry-specific terminology. Developers provide labeled training data containing examples of the entities they want the model to recognize. The resulting custom model can then process new text and identify those entity types. This capability is useful when standard named entity recognition does not provide sufficient coverage for specialized organizational or industry terminology.

Question 131

What is the primary purpose of embeddings in a retrieval-augmented generation solution?

1. Encrypt documents
2. Represent content as vectors
3. Convert speech to text
4. Generate database tables

Correct Answer: 2

Explanation

Embeddings represent text or other supported content as numerical vectors that capture aspects of semantic meaning. In a retrieval-augmented generation solution, documents and user queries can be converted into embeddings so their semantic similarity can be measured. A vector search system can then retrieve content whose embeddings are close to the query embedding. The retrieved information is supplied as context to a generative model, helping the model produce responses grounded in relevant source material. Embeddings therefore provide the numerical representation required for semantic vector retrieval rather than directly generating the final natural-language response.

Question 132

A Microsoft Foundry agent should retain relevant information from previous turns in an ongoing conversation. What concept supports this requirement?

1. Conversation state
2. Image segmentation
3. Document classification
4. Speech pronunciation

Correct Answer: 1

Explanation

Conversation state allows an agent application to maintain relevant information across multiple interactions. Without state, each user message could be processed independently, making it difficult for the agent to understand references to earlier requests or decisions. Maintaining appropriate conversation context allows the agent to handle follow-up questions more naturally and perform multi-step tasks. The application should still manage context carefully because excessive conversation history can increase token usage and latency. Sensitive information should also be handled according to security and privacy requirements. State management is therefore an important consideration when designing conversational agent experiences.

Question 133

Which Azure AI Search capability can return an exact count of matching documents when configured appropriately?

1. Faceting
2. Count retrieval
3. Speech recognition
4. Semantic captions

Correct Answer: 2

Explanation

Azure AI Search can be configured to return the count of documents that match a query. A result count is useful when applications need to display information such as the total number of matching records or determine whether a search produced enough results for a particular workflow. Counting is separate from features such as semantic ranking and vector similarity because it describes the number of matching documents rather than their relevance ordering. Developers should consider query performance when requesting counts, particularly in large indexes. Search response options should be selected according to the application’s functional and performance requirements.

Question 134

An AI application needs to store additional metadata such as document category and access level alongside text chunks in a vector index. Why is this useful?

1. It enables metadata filtering
2. It removes embeddings
3. It disables ranking
4. It converts vectors to images

Correct Answer: 1

Explanation

Metadata fields stored with vectorized content can support filtering and more controlled retrieval. For example, a document category can restrict results to policies, manuals, or technical guides, while an access-level field can help limit retrieval to content appropriate for a particular user or application context. Combining metadata filters with vector similarity can produce results that are both semantically relevant and contextually appropriate. Metadata is therefore an important part of index design in enterprise retrieval solutions. Developers should carefully define which fields are searchable, filterable, sortable, or facetable based on how the application will query them.

Question 135

Which Microsoft Foundry practice helps determine whether a generative AI application produces responses that meet predefined quality criteria?

1. Model evaluation
2. Image cropping
3. Audio compression
4. OCR preprocessing

Correct Answer: 1

Explanation

Model or application evaluation helps determine whether generative AI outputs meet predefined quality requirements. Evaluation can involve criteria such as relevance, groundedness, coherence, completeness, or other measures appropriate to the application. A development team can use representative test data and expected behaviors to compare different prompts, models, or application configurations. Evaluation is important because a response that sounds fluent is not necessarily accurate or appropriately grounded. Systematic testing provides evidence that can guide improvements before deployment and can also establish baselines for monitoring changes to application behavior over time.

Question 136

A developer is creating a custom text classification model for routing support requests. What type of training data is required?

1. Unlabeled images
2. Labeled text examples
3. Audio recordings
4. Video frames

Correct Answer: 2

Explanation

Custom text classification requires labeled examples that associate text with the categories the model should predict. For a support-routing application, examples might be labeled as Billing, Technical Support, Account Access, or another set of business-defined categories. The model learns patterns in the training examples and uses those patterns to classify new text. The quality and representativeness of the labeled dataset directly affect model performance. Training examples should cover the language, terminology, and variations expected in real user requests. Categories should also be clearly defined to reduce ambiguity during labeling and evaluation.

Question 137

An application must retrieve the most semantically similar documents from a small vector collection and prioritize exact similarity calculations over approximate performance. Which search approach can be used?

1. Exhaustive KNN
2. Semantic captions
3. Faceting
4. BM25 only

Correct Answer: 1

Explanation

Exhaustive KNN performs a nearest-neighbor comparison across the available vectors rather than relying on an approximate graph-based structure. This can provide highly accurate similarity results because the query is compared directly with the vector collection. The tradeoff is that exhaustive comparison can require more computation and may become less efficient as the dataset grows. Approximate methods such as HNSW are generally considered when search performance and scalability are more important. The appropriate approach depends on factors such as collection size, latency requirements, accuracy expectations, and available compute resources.

Question 138

A company wants to enrich documents with extracted entities before storing searchable information in an enterprise knowledge solution. Which Azure AI Search component is designed for enrichment processing?

1. Skillset
2. Query parser
3. Scoring profile
4. Search client

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 extraction, language processing, entity recognition, OCR-related processing, or other supported transformations. The enriched output can then be mapped into fields within the search index. This approach is useful for knowledge-mining scenarios where raw documents need additional processing before they become useful for search and retrieval. Skillsets can form part of an enrichment pipeline together with data sources, indexers, and indexes. The resulting structure can support richer enterprise search and retrieval applications.

Question 139

An agent must call a protected enterprise API without storing a user’s password in application code. Which approach is appropriate?

1. Hard-code credentials
2. Use managed identity
3. Disable authentication
4. Store passwords in prompts

Correct Answer: 2

Explanation

Managed identity provides an identity mechanism that can allow Azure resources and applications to authenticate to supported services without embedding long-lived credentials directly in application code. When configured correctly, the identity can be granted only the permissions required to access the target enterprise API or resource. This reduces the need to manage passwords or secrets within the application. Managed identity should be combined with appropriate role assignments, network controls, and authorization policies. The principle of least privilege remains important because authentication establishes identity, while authorization determines what actions that identity is permitted to perform.

Question 140

A generative AI application retrieves documents from an internal knowledge base but sometimes produces information that is not supported by those documents. Which improvement should be evaluated first?

1. Increase image resolution
2. Improve grounding and retrieval
3. Add speech synthesis
4. Disable vector search

Correct Answer: 2

Explanation

Improving grounding and retrieval can reduce unsupported responses by ensuring the generative model receives relevant and trustworthy context from the knowledge base. The retrieval stage should return appropriate documents or chunks for the user’s question, while the prompt should clearly instruct the model to use the supplied context and avoid unsupported claims. Developers can evaluate chunking, embeddings, search configuration, metadata filters, semantic ranking, and retrieval quality. Evaluation should also test cases where the answer is absent from the knowledge base. These measures help determine whether inaccurate responses originate from retrieval, prompting, or model behavior.