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

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.

Inside a Foundry project you can add an Azure OpenAI model straight from the model catalog. When you select GPT-4 Turbo and deploy it, Foundry automatically creates or links an Azure OpenAI resource and exposes it to the project through a managed connection secured with Microsoft Entra ID. Because the connection is handled by Foundry, developers authenticate with their Azure AD identities and no secret keys have to be shared. Creating a stand-alone Azure OpenAI resource, running the model in a container, or relying only on provider registration would either require manual key distribution or would not surface the model inside the project automatically.

Agents in Azure AI Foundry Agent Service can invoke external capabilities that are registered as tools (also called functions). By supplying an OpenAPI (or JSON schema) description for each REST endpoint and marking the invocation mode as automatic, the service allows the model to choose when a tool should be executed during the dialogue. Simply embedding endpoint details in the prompt (raw text) does not provide structured function metadata, a knowledge retrieval task targets indexed documents, and a Semantic Kernel planner skill would require explicit orchestration logic instead of autonomous tool selection.

In Azure AI Foundry, grounding a model with your own data follows a RAG pattern. After connecting the raw documents, you build a vector index to persist the embeddings that are created from the automatically chunked text. The index is implemented on top of Azure AI Search and is the resource the retrieval node in a prompt flow queries at runtime. Other data stores such as Cosmos DB, Redis, or AML Feature Store are not used by Foundry for vector retrieval in a RAG solution.

DALL-E 3 accepts three size values: 1024x1024 (square), 1024x1792 (portrait), and 1792x1024 (landscape). Among these, 1792x1024 is both landscape-oriented and the largest in total pixel count for that orientation. Therefore, setting size to 1792x1024 produces the largest landscape image. 1024x1792 is portrait, 512x512 is not supported by DALL-E 3, and 2048x1024 exceeds the model's current limits and will be rejected.

Azure AI Search indexers natively crack text-based formats such as PDF and can call built-in cognitive skills for images. However, audio and video enrichment relies on the Azure Video Indexer cognitive skill, which in turn requires that a Video Indexer resource be connected to your Azure subscription through an ARM-based account with an API key. Without creating and linking that account, the skill cannot extract spoken words or thumbnails, so MP4 and WAV files will fail. Enabling blob change feed, installing extensions, or deploying a Form Recognizer model does not meet this dependency.

The _{element with the alias attribute replaces the written text with an alternate pronunciation, so SQL forces "SQL" to be spoken as "sequel". To slow the delivery of that word, wrap the substitution in a element, which adjusts speaking rate. The required pause is achieved with a standalone element following the word. Other tags such as , , or do not simultaneously satisfy pronunciation substitution and rate control, or they lack the ability to add a timed pause.}

The Adult feature returns flags that indicate explicit or racy content. The Brands feature detects and identifies common brand logos such as Nike or Coca-Cola. The Description feature generates one or more concise captions of the image content. Requesting these three features satisfies all stated requirements without requesting unnecessary data, thereby reducing the response size. The other answer choices omit at least one required capability (adult content detection, brand recognition, or caption generation).

The groundedness metric compares each factual statement in the model's answer with the passages that the retriever supplied. If the text cannot be traced back to the provided context, the groundedness score drops. Therefore, a low groundedness score signals that portions of the answer are unsupported by the retrieved documents-an indication of hallucination. Fluency concerns grammatical quality, relevance measures how well the answer addresses the user's question, and semantic similarity compares the answer to a reference answer; none of these directly measure whether the answer is backed by the source passages.

Adding a system message that defines the model's persona, tone, and style, and includes an example of the desired output (a technique known as one-shot or few-shot learning), is the most effective way to control the brand voice. Lowering the temperature parameter makes the output more deterministic but does not teach it a specific style. A stop sequence is used to end generation at a specific point, not control tone. The presence_penalty parameter encourages the use of new words, which affects novelty, not the stylistic tone.

The Ogg16Khz16BitMonoOpus value instructs the Speech service to return 16-kHz, 16-bit, single-channel audio that is Opus-encoded and wrapped in an Ogg container. This combination provides good quality at a low bitrate and is broadly supported in web browsers. The RIFF and RAW PMC formats do not use Opus compression, so they produce larger files. Webm24Khz16BitMonoOpus uses the WebM container and a higher 24-kHz sample rate, resulting in higher bandwidth than required.

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 Read 3.2 API processes documents asynchronously. You start the analysis by calling the POST /read/analyze endpoint and supplying the PDF's SAS URL. The service immediately returns a 202 response that contains an Operation-Location header. Polling the URL in that header with GET requests returns status updates until the value becomes "succeeded". At that point the JSON payload includes a lines collection for each page that contains the extracted text and bounding boxes. Earlier Computer Vision OCR endpoints, synchronous receipt processing, and older recognizeText workflows do not meet the requirements for multipage PDFs processed with the Read 3.2 API.

In the JSON that the detectLanguage operation returns, each document object includes a detectedLanguage object. That object has a confidenceScore field whose value ranges from 0 to 1 and indicates how certain the service is that the identified language is correct. Statistics are only included when you request them and do not contain a language-confidence field. The property name is confidenceScore, not confidence, and the API no longer returns a detectedLanguages array, so reading the first element of such an array would fail.

The RecognizeReceipts (and other prebuilt) operations expose the optional query parameter includeTextDetails. Setting includeTextDetails=true instructs the service to include the full OCR hierarchy-pages, lines, words-and their bounding polygons in the analysis result. Changing the model version or content-type header does not alter the granularity of returned content. Switching to the Layout model would provide words and lines but would no longer return the receipt-specific key-value pairs required by the application.

The Analyze Image 3.2 endpoint supports multiple visual features in a single request, but it does not support optical character recognition. Printed or handwritten text must be extracted by a separate call to the Read API. Therefore you must first call Analyze Image with visualFeatures set to Objects and Tags, and then send a second call to the Read endpoint for the text. The other options are incorrect because:

• Read cannot be specified in the visualFeatures parameter.
• There is no separate features parameter that mixes Read with visualFeatures for version 3.2.
• The Description feature returns a caption and high-level tags only; it does not return object locations or extracted text.

In Azure AI Foundry the top-level construct is the hub. A hub is created at the subscription level and defines the catalog of foundation models, centrally managed policy settings, and shared infrastructure budgets for everything that is built beneath it. Individual teams then create one or more projects inside the hub; projects inherit the approved models but have their own prompt flows, experiment artifacts, and cost tracking. Therefore, creating a hub is the necessary first step before the separate team projects can be provisioned. Creating projects, resource groups, or OpenAI deployments can only occur after a hub exists, and none of those actions alone provides the shared model catalog required across teams.

When you choose the Unlabeled (formerly "template") training option, Azure AI Document Intelligence only needs sample documents that share a common layout. No label files or folder hierarchy is required, but the service enforces a minimum document count so that it can reliably infer the template. The container must therefore hold at least five sample documents of the same form type. Placing each document in its own folder, adding label JSON files, or including a model_data.json manifest are requirements for labeled training jobs, not for unlabeled ones. Filenames do not have to match any schema table.

Azure AI Language (formerly the Cognitive Service for Language) offers built-in capabilities for language detection, sentiment analysis, entity recognition, and extractive summarization. It also allows you to train and deploy custom multi-label text classification models, so all of the required tasks can be performed by one service. Azure AI Document Intelligence focuses on extracting structured data from documents, not general NLP. Azure AI Speech targets speech-to-text and related audio scenarios. Azure OpenAI Service provides powerful generative models but is not optimized for turnkey sentiment analysis, language detection, or custom text classification workflows that the Language service supplies out of the box.

Azure OpenAI fine-tuning jobs require a UTF-8-encoded JSON Lines (.jsonl) file. Each line contains one training example expressed as a messages array of role/content pairs (for example system, user, assistant). Every example must fit within the model's maximum context window-16 385 input tokens (plus up to 4 096 output tokens) for gpt-35-turbo. Other formats, such as a zipped JSON file, an Excel workbook, or separate prompt/completion text files, are rejected by Azure AI Studio.