Microsoft Azure AI Engineer Associate Practice Test (AI-102)

The Document Translation API is designed to work directly with Azure Blob Storage. A translation request supplies two shared access signature (SAS) URLs: one that points to the source container holding the original files and another that points to a target container where the translated copies will be written. Because the service pulls the documents from storage and pushes the results back without routing the files through your code, this approach keeps egress traffic low. Supplying raw document content-whether as base64, multipart/form-data, or a streamed WebSocket feed-would dramatically increase data movement and is not supported by the REST API for large-scale document translation.

The object detection results are returned in the "objects" collection. Each detected instance includes a "confidence" value ranging from 0-1 that indicates the model's probability for that specific object class. Reading the confidence of the object whose "object" field equals "person" lets you decide whether it meets the 0.8 threshold. The other listed properties represent scores for different features (category or tag detection) or do not exist, so they will not reflect the object-level probability you need.

The detailed, per-request information you require is emitted by the Azure AI Services platform as RequestLogs. By creating a diagnostic setting on the contoso-ai resource and selecting the RequestLogs (or AllLogs) category, you can stream those logs directly to an Azure Monitor Log Analytics workspace. Once the diagnostic setting is in place, every call-including its operation name, response code, and the key used-is recorded and becomes immediately queryable with Kusto queries. Exporting the subscription's activity log would only provide control-plane events, not request-level data. Connection Monitor tests and Application Insights dependency tracking would require additional configuration on every client and do not automatically capture the platform-level fields requested.

Adding a system message that explicitly instructs the assistant to output only the desired artifact narrows the response format. System messages establish the assistant's behavior across the entire conversation, so telling the model to return just the code in a single markdown block greatly reduces narration. Increasing temperature encourages more varied text, not less. A function-calling role could constrain the schema but requires additional parameter definitions and isn't necessary here, while asking for explanations invites even more non-code output.

Azure Video Indexer APIs require the caller to present the account's subscription key when requesting an access token. The key is supplied in the "Ocp-Apim-Subscription-Key" HTTP header. Other headers such as "Content-Type", "Accept", and "x-ms-client-request-id" are optional or used for different purposes and are not required for the access-token call.

Custom question answering stores and searches the content of your knowledge sources by using Azure Cognitive Search. Language Studio therefore requires an Azure Cognitive Search service (any pricing tier) to be linked to the Azure AI Language resource before the project-creation wizard can continue. The other resources-an Azure Storage account, Container Registry, or Bot Service-are not mandatory prerequisites for creating the project, although they might be used later for storage, container deployment, or bot integration.

The Adult feature runs the adult and racy content classifier, while the Tags feature returns a lightweight set of tags describing objects and concepts in the image. Together they satisfy the moderation and tagging requirements. Adding Objects, Read, Faces, or Description would trigger extra processing steps that are not needed, increasing latency and cost. Categories/Description do not perform adult detection, and any feature set that omits Adult would fail the moderation requirement.

To enable incremental crawling against Azure SQL Database tables, you include a dataChangeDetectionPolicy of type SqlIntegratedChangeTrackingPolicy in the indexer definition. This policy relies on SQL Server's native change tracking feature and allows the indexer to retrieve only the rows that have changed since its last successful execution. While a high-water-mark policy can be used with Azure SQL (typically for views), SqlIntegratedChangeTrackingPolicy is the recommended and most efficient method for tables. Schedule settings alone control when the indexer runs but do not handle delta detection. Resetting the indexer forces a full crawl rather than an incremental one.

Azure AI Foundry Agent Service stores each conversation's message history in a thread store. In a Standard agent setup you implement Bring-Your-Own (BYO) thread storage by connecting an Azure Cosmos DB for NoSQL account to the project. The service writes every user and assistant message to containers such as thread-message-store and can later retrieve them, so the agent can recall prior facts even when the user returns on another device. Azure Cache for Redis, Azure Storage, and Azure Service Bus do not integrate with the service for persistent thread storage.

Azure OpenAI Service exposes pretrained large language models such as GPT-4 and GPT-3.5 through REST and SDK endpoints, so no model training is required. The service also includes content filtering and other responsible-AI safeguards. Azure AI Language focuses on classic NLP tasks like entity recognition and summarization, not free-form text generation with large language models. Azure Machine Learning can host custom models but does not provide ready-to-use generative models or built-in filters. Azure Cognitive Search delivers indexing and search capabilities and cannot generate new text.

The service returns the redactedText field only when a masking redaction policy is applied. You enable masking by setting the redactionPolicy property (for example, policyKind = characterMask or entityMask). If you omit redactionPolicy or set it to noMask, the service returns the original text and the redactedText field is omitted. The domain, stringIndexType, and disableServiceLogs settings do not control whether redactedText is included.

Azure OpenAI automatically sends built-in platform metrics to Azure Monitor. In the Models - Usage category, the metric Total Tokens (REST name TotalTokens) counts the prompt and completion tokens consumed by the resource. Creating an Azure Monitor metric alert rule on that metric lets you trigger a notification when usage crosses a threshold such as 80 percent of your monthly quota-no additional instrumentation or diagnostic settings are required. Activity-log alerts, Advisor, and Log Analytics query alerts cannot directly evaluate running token totals unless the data is first streamed or collected separately, so they do not meet the requirement.

Azure OpenAI assigns each resource a default endpoint in the format https://.openai.azure.com. The custom sub-domain (contoso-opai) replaces and no region element is included in the hostname. Endpoints that add the Azure region, use cognitiveservices.azure.com, or append the resource name as a path are not valid default endpoints for an Azure OpenAI resource.

To identify specific individuals who appear in a video, Azure AI Video Indexer relies on a custom person model. Creating this model and populating it with reference images allows the service to match faces detected in the video with known people. Actions such as configuring CORS, brand detection or speaker name generation do not supply the visual reference data required for facial identification, so they cannot fulfil the requirement.

Azure AI Language provides pre-built REST and SDK operations for sentiment analysis, opinion mining, language detection, key-phrase extraction, named-entity recognition, and PII redaction. Because these capabilities are available out of the box, no model training or hosting is required. Azure OpenAI could achieve similar results but would need custom prompt engineering and carries higher cost. Azure AI Content Safety focuses on harmful content detection, not broad linguistic analysis. Azure AI Document Intelligence specializes in extracting structured data from scanned documents and forms rather than analyzing unstructured review text.

Updating the existing deployment to point to the newer model version keeps the deployment name unchanged, so all client applications continue to call the same endpoint and API path. Because the change is confined to the deployment configuration, you can roll forward or roll back at any time by editing the deployment again, without touching application code or provisioning a new resource. Deleting the deployment first causes outage, while creating a separate deployment or resource would require every caller to update its deployment name or endpoint, violating the no-code-change requirement.

A private Azure OpenAI deployment is surfaced to Foundry projects by adding the existing Azure OpenAI resource. To hold embeddings that can be refreshed, you create a managed Vector Index resource inside the project, which automatically generates the secure connection needed by prompt flows. A Key Vault alone, Cosmos DB, or App Configuration do not provide integrated vector-index functionality.

In Azure AI Foundry the first step of a Retrieval-Augmented Generation (RAG) solution is to make the unstructured data searchable. You do this by creating a vectorized Azure AI Search index that stores both the text chunks and their embeddings. The retrieval step in a prompt flow can then query this index and pass the most relevant chunks to the GPT-35-Turbo prompt. If you skip the indexing phase-or use a regular (keyword-only) index-semantic search and vector similarity ranking will not work, and the model cannot be properly grounded. Uploading the PDFs alone or adding them to the project's data assets does not provide the low-latency semantic search that RAG requires.

CLU always reserves a portion of the data (20 percent by default) as a blind test set. Because the CancelAppointment intent has fewer than the recommended 15 training examples, only a handful of test utterances are available after the split. With so few samples, a single misclassification can drop recall to 0 percent or raise it sharply in another run, making the reported metrics volatile. Adding more labeled utterances-especially for intents with fewer than 15 examples-will give the model more reliable training data and provide a larger test set, stabilizing the evaluation scores. The other options are incorrect because CLU does not switch to k-fold cross-validation, entity labels are not required for intent metrics, and standard training metrics are deterministic once the data volume is adequate.

Azure OpenAI provides a built-in content safety system that can be configured per deployment. By enabling the content filtering policy and setting the Violence and Sexual categories to the block severity threshold, the runtime will automatically refuse or safe-complete requests that include disallowed content while still allowing benign requests. Web application firewalls and generic text analytics profanity filters do not inspect model generations deeply enough for Responsible AI compliance, and a private endpoint only secures network access-it does not moderate content.