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

Incremental indexing is enabled by adding a HighWaterMarkChangeDetectionPolicy to the indexer. You specify the name of a monotonically increasing column (such as LastModifiedDate) and the indexer automatically stores the highest value it has seen. On the next run it requests only rows where the column value is greater than that stored watermark, thereby importing just new or updated data. Changing the schedule merely determines when runs occur, batchSize controls how many items are sent per batch, and a SoftDeleteColumnDeletionDetectionPolicy is used to detect deletions, not updates.

A custom question answering project is always linked to a single language locale. To provide answers in more than one language, you create a separate project for each language (for example, one project with locale en-us and another with locale es-es). At run time your client application detects the user's language or inspects the Accept-Language header, then routes the request to the endpoint of the matching project. Adding multiple languages to one project, translating queries on the fly, or relying on cross-language embedding are not supported ways to build a production multi-language knowledge base in Azure AI Language.

The gpt-35-turbo-16k model offers a 16 000-token context window-enough for the 14 000-token inputs-and natively supports the Azure OpenAI function calling feature. It is significantly less expensive than any GPT-4 variant, making it the most cost-effective choice that still fulfils both the context-length and function-calling requirements.

GPT-4-32k also meets the technical requirements but costs considerably more. text-davinci-003 is a completions model with only a 4 000-token limit and no function-calling capability. text-embedding-ada-002 is an embeddings model, not a chat/completions model, so it cannot generate summaries or call functions.

Phi-2 is a 2.7-billion-parameter open-weight language model published by Microsoft. Because the weights can be downloaded from the Azure AI model catalog, the model can be executed completely offline and even on a single consumer-class GPU, making it suitable for edge scenarios that lack network connectivity. GPT-4 and GPT-3.5-Turbo are only available as hosted Azure OpenAI endpoints, so they require cloud access and cannot be run locally. DALL-E 3 is an image generation model, not a text-generation model, and also relies on a hosted service.

Opinion mining is an optional extension of the Analyze Sentiment operation. You enable it by adding the query parameter "opinionMining=true" (or the same property in the SDK options). When this flag is present, the service extracts aspect-opinion pairs and returns sentiment and confidence scores for each pair in the response. Parameters such as "showStats" or "stringIndexType" control statistics and text index calculation but do not activate opinion mining, and "includeOpinionMining" is not a recognized parameter.

The free F0 tier is limited to 5,000 text records per month, so it cannot support the current volume. Moving the resource to another region may offer minor price differences but does not provide the substantial discounts available for high-volume, predictable workloads. Disabling public network access is a security measure and has no impact on transaction billing. Creating an Azure budget helps monitor spending and sends alerts, but it does not inherently reduce charges. Purchasing a monthly commitment tier for Azure AI Language provides a significant discount compared to pay-as-you-go pricing for predictable usage volumes, making it the most effective way to lower costs without architectural changes.

The Azure Function can acquire an Azure AD access token for Azure OpenAI through its system-assigned managed identity. When you instantiate OpenAIClient with an Azure AD token credential (for example, DefaultAzureCredential), the SDK handles token acquisition transparently. For the token to be accepted, the managed identity must be granted the Cognitive Services OpenAI User role on the Azure OpenAI resource. Passing the managed identity's object ID as an API key (or retrieving the key from Key Vault) would still rely on key-based authentication, which the security team disallows. Assigning the Reader role or using AzureKeyCredential does not enable token-based authentication.

When you create an agent with the Azure AI Foundry Agent Service you must identify which large language model the runtime will call. The CLI therefore requires that the JSON specification include an underlying chat model reference (for example, gpt-35-turbo or gpt-4o). Without the model element the creation request is rejected because the service cannot determine where to send user prompts. Properties such as an instruction set, a memory store, or an Azure AD-backed managed identity are optional at creation time and can be added or updated later.

Azure AI Document Intelligence is purpose-built to extract structured data such as key-value pairs, tables, and line items from forms and business documents, including invoices. It can return the identified fields as JSON with little or no custom model training through its pre-built Invoice model. The Read OCR capability in Azure AI Vision only returns raw text; you would still need custom logic to locate and label invoice fields. Named Entity Recognition in Azure AI Language works on already extracted plain text and is designed to find entities such as people or locations, not document-layout-based fields. Azure Cognitive Search focuses on indexing and searching content; while it can orchestrate enrichment skills, it is not the dedicated service for high-accuracy field extraction from invoices.

The Analyze Image endpoint lets you request multiple analysis operations at once by supplying a comma-separated list to the visualFeatures parameter. To receive both bounding boxes (object detection) and descriptive image tags in the same response, you must include both Objects and Tags in that list. Supplying only one of the feature names-or unrelated features such as Categories-will omit the other required data, and passing multiple values in separate requests defeats the requirement to use a single call.

Running the Custom Vision runtime container on an edge device keeps all inference on-premises, eliminates the need for constant Internet connectivity, and provides sub-second latency. Publishing to the cloud prediction endpoint or to an AML managed endpoint still requires outbound connectivity and adds network latency. Azure OpenAI cannot host vision models, so that option is not applicable.

Azure Monitor workbooks can only query data that has been sent to a Log Analytics workspace. An Azure OpenAI resource emits content-filter events through the ContentFiltering log category. By creating a diagnostic setting on the resource, selecting the ContentFiltering category, and routing the logs to a Log Analytics workspace, you persist every allow, block, or safe-complete decision together with its category and severity. Networking settings, Key Vault diagnostic settings, or Azure Policy assignments do not collect or store the required filter telemetry, so they will not populate the workbook.

Keyword spotting is performed completely on the client device. In the Speech SDK you create a KeywordRecognitionModel from the .table file that Speech Studio produces for the custom wake word, and then pass that model to a KeywordRecognizer (or a SpeechRecognizer started with StartKeywordRecognitionAsync) so that audio is evaluated locally until the keyword is detected. The other options would either rely on cloud processing (IntentRecognizer, DialogServiceConnector) or use APIs that do not accept a keyword model (.table files cannot be loaded as grammars or language-detection resources). Therefore, only the combination of KeywordRecognizer and KeywordRecognitionModel satisfies the offline wake-word requirement.

The QuestionAnsweringClient constructor requires two parameters that allow the SDK to authenticate and route the request: the HTTPS endpoint URI of the Azure AI Language resource and an AzureKeyCredential that contains one of the resource keys. The project name and deployment name are supplied later when you call the client's QueryKnowledgeBaseAsync method, while the resource's region is implicit in the endpoint URL. The subscription ID and authoring key are not used by the runtime SDK.

DALL-E 3 offered through Azure OpenAI is the only catalog model that both embeds C2PA metadata in every generated image and is integrated with Microsoft's image content-safety filters. Stable Diffusion XL can reach the required resolution but lacks automatic C2PA credentials and does not use the built-in Azure OpenAI safety pipeline. GPT-4 Turbo and Llama 2-70B-Chat are text-only and cannot produce images.

The Azure AI Foundry SDK can invoke a prompt flow only after the flow is deployed as an online endpoint. Publishing or validating a flow makes it available for manual tests in the portal but does not provision the fully managed inference endpoint that the SDK targets. Selecting Deploy > Online in the portal (or calling the corresponding deploy_online method in the SDK) creates the endpoint and activates the REST and SDK surface so the invoke operation resolves. Other actions such as exporting the flow, promoting it to staging, or regenerating credentials do not create an endpoint and therefore would still result in a 404 error.

An Azure AI Search indexer can connect to the blob container and run a skillset that includes the built-in OCR skill (backed by Azure AI Vision) and the Key Phrase Extraction skill (backed by Azure AI Language). This automatically reads text in JPG and PDF files and generates topic keywords, then writes the enriched output to the target search index without additional code. Azure AI Video Indexer indexes MP4 files, producing speech-to-text transcripts, detected topics, and other insights. Its REST API or sample power-skill can export those insights directly into the same search index. Together, these two services satisfy all requirements while requiring minimal orchestration.

The Azure SDKs emit OpenTelemetry spans only when a tracing implementation is attached. Installing and importing the azure-core-tracing-opentelemetry package automatically patches all Azure SDK client libraries-including azure-ai-openai-to create detailed spans for each service call. These spans include service-side latency (the value of the "request_process_time" response header), token usage (values in the "x-ms-usage-tokens-prompt" and "x-ms-usage-tokens-completion" headers), and the request/response bodies up to 8 kB. Because your project already initialises an OpenTelemetry exporter for Application Insights, adding this package is the only required step. Other options either capture only logs, require manual span creation, or send data to a different monitoring solution, so they do not meet the requirements.

The Speech SDK selects a custom speech-to-text model when the client specifies the model's endpoint ID. This is done by assigning the deployment (endpoint) ID to the SpeechConfig.EndpointId property and then passing that SpeechConfig object to the SpeechRecognizer constructor. Modifying SpeechRecognitionLanguage only chooses a base language; adding the ID to the subscription key or to event handlers has no effect on model selection, so those approaches will still route audio to the standard model.

Using a user-assigned managed identity lets the AKS workload obtain an OAuth 2.0 access token from Azure AD without storing any secrets. After you assign the identity to the AKS cluster and grant it the Cognitive Services User role on the Azure AI Foundry resource, the application can request tokens and call the endpoint. Keys, SAS tokens, or IP firewalls do not meet the requirement to avoid static credentials or provide automatic rotation.