Microsoft Azure AI Engineer Exam Dumps and Braindumps

The correct answer is Edit the AI Search solution by adding fields to the index and updating the indexer and custom skillset to map the model output.

This option is correct because enrichments produced by a deployed classification model must be mapped into the search index schema and then written by the indexer. Adding fields to the index gives you a place to store the genre tags and updating the indexer and custom skillset ensures the model output is transformed and routed into those fields so the enrichments persist and become queryable facets.

Create an Azure Function app to invoke the deployed classification endpoint is incorrect because simply invoking the endpoint does not by itself write the classification results into the AI Search index. You would still need to integrate the invocation into the enrichment pipeline and map outputs to index fields.

Train a custom text classification model in Language Studio is incorrect because training the model alone does not persist enrichments in the search index. The model must be connected to the search enrichment pipeline and its outputs must be mapped into index fields.

Train the classification model and also deploy an Azure Function app to handle requests is incorrect because deploying a function to call the model still does not complete the end to end flow unless you update the index schema and configure the indexer or a custom skill to map the responses into index fields.

Only update the index schema without creating a custom skill or configuring indexer output mappings is incorrect because changing the schema without creating the enrichment step and the output mappings leaves the new fields empty and the classification tags will not be persisted in the index.

Focus on end to end data flow when you answer these questions. Ask whether the model output is actually written into index fields and whether the indexer or skillset is configured to map and persist those outputs.

The correct option is Chat Completions.

Chat Completions is built for multi turn conversational agents because it accepts a structured sequence of messages with roles such as system, user, and assistant and it preserves conversation context across turns so the model can respond as part of an ongoing dialogue.

Chat Completions also supports conveniences that make building agents easier such as streaming responses and clearer role handling so you do not need to manually concatenate long prompts to represent prior turns.

Azure Bot Service is a platform for building, hosting, and connecting bots to channels and it is not itself the model inference API endpoint used to generate chat responses, so it is not the correct API endpoint to call to perform multi turn model generation.

Completions is a single turn text completion endpoint that is intended for one off prompts and it lacks the structured messages format with explicit roles that simplifies multi turn conversation, so it is not the ideal choice for building a multi turn conversational agent.

When the question asks about multi turn conversation look for the API that accepts a messages structure with roles. Use Chat Completions for ongoing dialogue and reserve Completions for simpler single turn text generation.

WaterfallDialog is the correct option for managing a strictly ordered sequence of setup steps in a device onboarding workflow.

WaterfallDialog models a linear sequence of steps where each step executes in order and can prompt the user or perform processing before advancing to the next step. This predictable step by step flow makes it well suited for onboarding scenarios where each setup action must complete before the next one begins.

AdaptiveDialog is not correct because it is built for flexible, rule driven and event driven conversations that can change flow dynamically rather than enforcing a fixed ordered sequence.

ComponentDialog is not correct because it is a container for composing and reusing dialogs and it does not by itself provide the linear step sequencing required for a strict onboarding flow.

ActionDialog is not correct because it focuses on invoking or managing discrete actions and is not the pattern used to orchestrate a simple, ordered series of setup steps.

When the question emphasizes a strictly ordered or step by step user journey choose the dialog type built for linear sequencing such as WaterfallDialog.

The correct option is 96%.

Precision is the number of true positives divided by the number of predicted positives. The model made 50 predictions and 48 were correct, so the precision is 48 divided by 50 which equals 96 percent and matches 96%.

0.96 is incorrect as an option in this question even though it represents the same numeric value as 96 percent. The available correct choice was presented as a percent so the percent form 96% is the answer to report.

0.92 is wrong because it would correspond to 46 correct out of 50 and that does not match the given 48 correct predictions.

48 out of 50 simply states the raw count of correct predictions but it does not present the precision in the requested reported format. The correct reported precision is 96%.

When asked for precision compute true positives divided by predicted positives and then match the format shown in the answer choices.

The correct option is JSON.

The Entity Recognition API requires a JSON request body with a documents array where each document contains an id and a text field and an optional language field. The service expects the text for each record to be a string and the maximum permitted size per record is 45,000 characters as measured by String.Length.

YAML is incorrect because the API does not accept YAML payloads and requires JSON formatted requests according to the service specifications.

Protocol Buffers is incorrect because Azure Language REST endpoints use JSON over HTTP and do not use binary Protobuf messages for their public REST API.

XML is incorrect because the service expects JSON and not XML payloads for the Entity Recognition endpoints.

Plain text is incorrect because you must send structured JSON with document entries and simple plain text bodies will not meet the API request schema.

Focus on the request examples and the request body schema and remember that Azure Language REST APIs expect JSON with a documents array and check field length limits when they are mentioned.

The correct answer is Add the Direct Line Speech channel to chatAgent2.

Add the Direct Line Speech channel to chatAgent2 extends the agent by adding real time speech input and output which makes the agent accessible to users who rely on voice or who prefer speaking over typing. Integrating the speech channel brings speech to text and text to speech capabilities which supports multimodal interactions and helps the agent meet inclusive interaction goals.

Direct Line Speech also ties into accessibility and inclusivity because it lowers barriers for people with mobility or vision challenges and it supports a broader range of user preferences. This aligns with responsible AI principles that emphasize accessibility and equitable access to services.

Enable active learning on nlm2 is incorrect because active learning focuses on collecting data and improving model training and it does not by itself add new input modalities or accessibility features that make interactions more inclusive.

Require Azure Active Directory sign in for chatAgent2 is incorrect because enforcing Azure AD sign in improves security but can create access barriers and it does not provide additional speech or accessibility channels that support more inclusive interactions.

When a question asks about making interactions more inclusive look for options that add or improve input and output modalities. Adding speech or other accessibility channels is often the right choice rather than changing training flags or requiring stricter authentication.

The correct option is Provision a private endpoint for the search instance.

A private endpoint uses Azure Private Link to assign a private IP address in your virtual network and it ensures traffic to the SearchPrime instance flows over the Microsoft backbone rather than the public internet. This means the search service is not reachable from public IP addresses and you can enforce network level controls inside your VNet.

Using a private endpoint is the way to fully remove public exposure because it places the service on a private address and integrates with private DNS and network security policies.

Apply an IP access control list to the search service is not sufficient because IP ACLs still leave the service with a public endpoint. Firewall rules restrict which public IPs can connect but they do not remove the public internet surface.

Enable virtual network service endpoints for the search resource is incorrect because Azure Cognitive Search does not use service endpoints to become privately reachable. Service endpoints do not place the service inside your VNet and they are not the supported method to eliminate public access for Azure Search.

When you need to remove public internet access prefer solutions that use Private Link or private endpoints because network ACLs still rely on a public endpoint and do not fully eliminate exposure.

The correct answer is Classification.

The Classification field contains category entries and numeric score values that represent the model confidence for each content category such as sexual content. These score values let you threshold and interpret how likely the text belongs to a given category.

PII is focused on detecting personally identifiable information like names, email addresses, and phone numbers and it does not return per-category confidence scores for content categories.

Terms lists words or phrases that matched the moderation term lists and it reports matched text and offsets but it does not provide numeric confidence scores for category classification.

When scanning API responses look for fields named Score or Classification to find numeric confidence values that indicate how strongly a piece of content matches a category.

Create the search resource and build the index in Azure AI Studio is the correct option for adding content when using Azure OpenAI on your data.

Create the search resource and build the index in Azure AI Studio is correct because Azure AI Studio provides an integrated flow to provision an Azure Cognitive Search resource, define the index schema, and ingest documents so that the search index can be used with the Azure OpenAI on your data features. The studio centralizes connector setup, indexing options, and semantic and vector search settings so you have a single place to prepare content for retrieval augmented generation scenarios.

Use the Azure Cognitive Search portal to build and populate an index is not the recommended answer in this context because the question asks about the workflow when using the Azure OpenAI on your data experience in Azure AI Studio. The standalone Azure Cognitive Search portal can be used to manage search resources, but the integrated experience in Azure AI Studio is the recommended path for building the index for Azure OpenAI scenarios.

Choose any available data source option in Azure OpenAI on your data is incorrect because you must first create and populate a search index that the service uses for retrieval. You cannot simply pick any generic data source from within the Azure OpenAI on your data UI and expect the retrieval layer to be configured without building the search resource and index.

Use Azure AI Studio to connect to OneDrive Dropbox Box or Google Drive resources is incorrect as the single recommended step. While connectors may be available or evolving, the key required action is to create the search resource and build the index so that the data is ingested and searchable by the Azure OpenAI on your data capability.

When a question asks for the recommended method look for the option that creates the search resource and index in Azure AI Studio because that is the integrated, supported workflow for Azure OpenAI on your data.

Switch to General (compact) then retrain then export the trained model for offline use is correct.

The compact domain produces models that are optimized for size and for edge or offline deployment. After you change the project to the General (compact) domain you must retrain so that the model weights are produced for that domain. Once training completes you can export the trained model in an offline format such as ONNX or a TensorFlow variant for deployment on an isolated network.

Create a new classification project then switch to General (compact) then export is incorrect because a classification project is for image classification and not for object detection. Creating a classification project will not produce the bounding box detection outputs that an object detection deployment requires.

Export the existing model then change domain to General (compact) then retrain is incorrect because exporting first does not convert the model to the compact domain. The domain must be set and the model retrained under that domain before you export the artifact for offline use.

Remember to set the project domain to a compact option and then retrain before you export. The export will only produce a proper offline model after training completes under the compact domain.

The correct answer is a dedicated Cognitive Services instance on the S0 pricing tier.

The a dedicated Cognitive Services instance on the S0 pricing tier gives you a paid, multi API endpoint that supports both OCR and text analytics and it integrates cleanly with an Azure Cognitive Search skillset. It provides higher throughput and predictable quotas for production scale ingestion of many documents which helps keep overall costs lower compared with unpredictable per-call or throttled free limits when you must process a large batch of scanned files.

a free Cognitive Services resource with limited enrichments is incorrect because the free offering has restricted quotas and capabilities and it is not suitable for production scale enrichment pipelines. The free tier often limits the types of enrichments available and it can be throttled which would slow or block a large indexing job.

an Azure Computer Vision resource is incorrect because Computer Vision mainly provides image focused capabilities such as OCR and visual analysis and it does not cover the full set of text analytics enrichments like language detection and entity extraction that a combined cognitive services instance provides. Using only Computer Vision would require additional services to get full text analytics.

an Azure Machine Learning Designer pipeline is incorrect because Designer pipelines are not directly attachable as a built in skill to Azure Cognitive Search. You could expose a custom web API backed by a model as a custom skill but that adds deployment complexity and cost and it is not the simplest or most cost effective choice for standard OCR and text analytics.

When a skillset needs both OCR and text analytics choose a paid multi API Cognitive Services instance for predictable quotas and supported enrichments and avoid the limited free tier when processing large document sets. Consider using a custom skill only if you need a model that is not available in Cognitive Services.

The correct answer is Compress WAV files with matching transcript into a ZIP archive.

Speech Studio expects training data to be uploaded as a packaged dataset so that audio files and their transcripts remain paired and the importer can validate formats and filenames. A ZIP archive that contains WAV audio and plain text transcripts with matching names is the supported and simplest way to import bulk samples for training or evaluation.

Store files in Azure Blob Storage and link them is incorrect because the exam scenario asks about how to package samples for import into Speech Studio and the service expects a defined archive format for bulk import. While you can store assets in Blob Storage for other workflows you still need to prepare a proper ZIP dataset for import into Speech Studio.

Upload individual FLAC files and attach a DOCX transcript is incorrect because FLAC and DOCX are not the typical bulk import formats for Speech Studio. Transcripts should be plain text files aligned to audio filenames and audio is normally provided as WAV in the supported encoding so the importer can pair and validate them.

When a question asks about importing data for Speech Studio remember that the safest choice is usually a ZIP of WAV files with plain text transcripts that have matching filenames.

The correct answer is 1.2 GB. This is the largest single file size that can be uploaded for agents or for fine tuning in Orion AI Lab.

1.2 GB is enforced as the single file upload limit so uploads remain reliable and fit within the platform resource constraints. If you split a dataset into multiple files you can upload more total data but each file must stay at or below the 1.2 GB limit.

640 MB is incorrect because it is smaller than the actual single file limit and would understate the allowable upload size.

320 MB is incorrect because it is well below the supported single file size and does not match the platform limit.

240 GB is incorrect because it greatly exceeds the single file upload limit and is not a realistic value for a single file upload to this service.

Carefully note whether the question asks about a single file limit or a total dataset limit and always compare the numeric value and its units before selecting an option.

The correct answer is Speech.

The Speech service is designed to process audio and it provides real time speech to text, voice activity detection, and speaker diarization which together let you detect when a learner is speaking in an audio stream.

Video Indexer is incorrect because it is focused on analyzing and indexing video content and while it can extract and transcribe audio tracks it is a higher level video analysis service rather than the core audio speech service for detecting live speech.

Face is incorrect because it analyzes images and facial attributes and it does not process audio so it cannot detect speech from an audio stream.

When you see a question about processing live or streamed audio look for services related to Speech and save image or video services for visual tasks.

Add logo examples to the Video Analyzer Brands model and include Bing suggested entries is correct.

Add logo examples to the Video Analyzer Brands model and include Bing suggested entries is the right approach because the Brands capability is built to find logos in video content and it accepts example images to expand and refine detection. By supplying labeled logo examples you teach the Video Analyzer Brands model to recognize custom brands and the Bing suggested entries help match common variations and known brand metadata so detection is more robust across many videos.

Add logo examples to the Video Analyzer Brands model and include Bing suggested entries also keeps the workflow inside the video analytics pipeline so you do not have to extract frames, host a separate image classifier, and manage a separate training and inference flow for each video. You can incrementally add new brands to the model so the detection set grows as new logos are required.

Video Analyzer for Media may appear in older materials and you should be aware that Microsoft has evolved its video offerings so newer exams and documentation might reference Video Indexer or Azure Media Services for equivalent capabilities.

Train an Azure Custom Vision logo classifier is not the best single answer because Custom Vision is an image classification service and it requires you to extract frames and manage a separate model and inference pipeline for video. It can work for logos but it does not provide the built in video brand detection and Bing suggestion integration that the Brands model offers.

Use Computer Vision to extract key frames from videos is incorrect because Computer Vision focuses on image analysis tasks and it does not provide an integrated brand/logo recognition model or the Bing suggested brand entries. Extracting frames with Computer Vision would be an extra manual step and you would still need a brand matching solution.

Embed Video Analyzer widgets in a custom portal that stores reference brand images is wrong because a portal and reference images alone do not perform automated brand detection across video files at scale. That approach is a UI or storage pattern and it does not replace a trained brands model that runs detection on video content.

When a question asks about detecting logos in video look for answers that mention a built in brand or logo model and the ability to add examples. That usually beats building separate image classification pipelines. Also watch for service name changes because Azure video offerings are often updated.

Match only the exact spoken text is correct. This choice is appropriate when you need to accept only specific words or phrases exactly as spoken and you do not want any machine learning inference to modify or generalize the result.

Pattern matching is deterministic and works well when the set of valid utterances is known in advance and must be matched literally. In those cases you can apply simple rules or regular expressions to the speech transcript to confirm exact phrases rather than relying on probabilistic models that try to interpret intent or normalize entities.

Prebuilt entity recognition is incorrect because prebuilt extractors use trained models to identify common entities and they will attempt to generalize from varied language. That behavior makes them unsuitable if the requirement is to match only the exact spoken text.

Machine learned entity recognition is incorrect because machine learned extractors are probabilistic and designed to infer entities and intents from diverse utterances. They are valuable for handling variability but they do not guarantee exact literal matches.

When the question emphasizes words like exact or only prefer rule based pattern matching. Machine learned or prebuilt recognizers are better when you need flexibility rather than literal matching.

The correct option is No.

Enabling a public endpoint on the search service and then limiting access with an IP firewall does restrict which clients can connect but it still exposes the service on a public address and can cause traffic to traverse the public network rather than remaining entirely on the Azure backbone. Because the requirement is that traffic must not leave the Azure backbone this approach does not satisfy the requirement.

To meet the requirement you should use an Azure Private Endpoint or Private Link for Azure Cognitive Search so the service receives a private IP address in the virtual network and traffic flows over the Microsoft network without going to the public internet.

Yes is incorrect because using an IP firewall on a public endpoint controls access but does not change the traffic path and therefore does not guarantee that traffic stays on the Azure backbone.

When a question asks to keep traffic on the Azure backbone prefer using an Azure Private Endpoint or Private Link rather than a public endpoint with IP firewall rules.

The correct answers are URLs, Numbered and bulleted lists, and Formatted text.

Azure AI Language question answering can ingest content referenced by URLs so the service can fetch web pages and extract text during import. The importer also preserves structure such as Numbered and bulleted lists which helps the model understand ordered and unordered items. The system also retains Formatted text from supported documents so headings and paragraphs remain available for indexing and answer generation.

Images is incorrect because the question answering import focuses on textual content and structured text from files and web pages. Images are not extracted as text by the importer unless they are processed first with OCR or converted into a supported document format that contains selectable text.

When studying ingestion features remember that question answering expects textual inputs so images usually require OCR or conversion to a supported text format before they count as imported content.

The correct option is prebuilt receipt model.

The prebuilt receipt model is a turnkey solution that is trained to recognize common receipt fields such as merchant name purchase date time of purchase taxes paid and final amount directly from scanned images so it delivers structured output without custom training.

Because the prebuilt receipt model returns named fields and handles variations in layout you only need to call the API and map the results into your application which makes it the least implementation effort for this use case.

custom neural model is incorrect because building a custom neural model requires collecting labeled receipts training the model and maintaining it which increases development effort.

prebuilt Read model is incorrect because the Read model provides OCR and raw text extraction rather than structured receipt fields so you would still need to parse and identify merchant totals and taxes yourself.

custom template model is incorrect because template or custom models need samples and training to extract specific fields and they are not as immediately ready to extract receipt fields as the prebuilt receipt model.

When the question asks for the least implementation effort look for a prebuilt model that returns structured fields for the document type instead of raw OCR or models that require training.

The correct answer is Upload a ZIP archive containing individual WAV files and a matching plain text transcript.

Upload a ZIP archive containing individual WAV files and a matching plain text transcript is correct because supervised training requires each audio sample to be paired with its exact transcript so the model can learn sound to text mappings. A ZIP lets you submit many files in a single package and individual WAV files provide a widely supported, lossless audio format that preserves quality. Plain text transcripts are easy for training systems to parse and to match to each audio file by filename.

Provide individual FLAC files and a transcript saved as a Word document is incorrect because while FLAC is a lossless audio format some training pipelines expect standard WAV files and a Word document includes formatting that prevents reliable automatic parsing and pairing with audio files. Training services typically require plain text transcripts rather than formatted documents.

Upload each recording as a single merged WMA file without transcripts is incorrect because merging recordings prevents per-sample alignment and supervised learning. WMA is less commonly supported than WAV and omitting transcripts means there is no ground truth for the model to learn from.

Keep each audio file separate and provide plain text transcripts with matching filenames so the training system can automatically pair and validate the data.

The correct answer is Computer Vision analyze images endpoint.

The Computer Vision analyze images endpoint provides general image analysis features and it specifically returns information about the image type so it can indicate whether an image is a photograph a drawing or clipart. This endpoint gives general labels tags and an imageType result that makes it suitable for broad classification of many images without creating a custom model.

Custom Vision image classification is not the best choice because it is meant for training custom classifiers on user supplied labels and it requires you to create and train a model when the built in general image type detection already covers photograph drawing and clipart scenarios.

Computer Vision object detection is incorrect because object detection focuses on finding and localizing objects with bounding boxes and classes and it does not provide the general image type classification that the analyze images endpoint returns.

Custom Vision object detection is wrong because it is used to train custom object detectors that return bounding boxes and custom classes and it is not intended for the generic photograph drawing or clipart classification that the analyze images endpoint supports.

When a question asks about recognizing whether an image is a photograph drawing or clipart look for the service that returns an imageType result. If you need user specific categories or bounding boxes then consider Custom Vision or object detection respectively.

Create a search data source pointing to the blob container is correct.

Creating a search data source in Azure AI Search registers the blob container as a content source and provides the service with the connection information and access method it needs. The search indexer then uses that data source to crawl the container and extract content from Word documents so the search index can be populated.

Upload an index JSON file to the blob container is incorrect because index definitions are created and stored in the search service rather than uploaded into the blob container as a JSON file. You do not make the index available to Azure AI Search by placing a JSON file in the container.

Grant a managed identity the Storage Blob Data Reader role is also incorrect as the sole required step because assigning permissions alone does not register the container with the search service. While a managed identity or other credentials can be used for authentication, you must still create the search data source and indexer so Azure AI Search knows where to find and how to pull the documents.

When a question asks how to let Azure Search read files from storage think about configuring a data source and an indexer rather than uploading files that represent the index or only setting permissions.

The correct answer is Add each target language to the translation configuration by calling the configuration method on the SpeechTranslationConfig object.

You configure multiple output languages on the SpeechTranslationConfig object before creating the recognizer. When you add several target languages the SDK will request translations for each configured language in a single recognition pass so you get translations for all targets from one audio input. This approach is efficient and keeps the client code simple because the recognizer returns translation results for every configured target language.

Duplicate the application module for every language and update its language settings is wrong because duplicating modules is unnecessary and hard to maintain when the SDK already supports multiple target languages through configuration.

Append an extra language pair parameter to the translation function call is wrong because the SDK expects target languages to be set up in the translation configuration rather than passed as ad hoc extra parameters on each function call.

Implement separate translation functions so each routine handles a single language pair is wrong because creating separate routines adds complexity and overhead and is not required when the SpeechTranslationConfig object can produce translations for multiple languages in one workflow.

When a question mentions producing multiple output languages think about SDK configuration objects like SpeechTranslationConfig which usually let you add several targets up front instead of duplicating code.

The correct answer is Both ingestion and retention are billed.

Azure Monitor Log Analytics charges for the volume of data that is ingested into the workspace and it also charges for keeping that data for longer periods. Ingestion fees cover the bytes sent into Log Analytics and retention fees apply when you store data beyond the included free retention period or when you choose to retain data longer than the default settings.

You can manage costs by adjusting retention settings or by choosing commitment tiers and capacity reservations that change how ingestion and long term storage are billed.

Only ingestion is billed is incorrect because retention storage is also charged when data is kept beyond the included retention window or when longer retention is selected.

Neither ingestion nor retention is billed is incorrect because Azure Monitor Log Analytics does apply charges for both ingesting data and for retaining data for extended periods.

When a pricing question mentions log or monitoring services look for both the terms ingestion and retention in the answers. Cloud logging providers commonly bill for bytes ingested and for longer term storage so answers that include both are often correct.

The correct option is Add follow up prompts.

Add follow up prompts configures the question answering project to present explicit next step choices and it lets the system link answers across turns so the agent can continue a multi turn dialogue and maintain context. This approach enables the agent to ask clarifying or guided follow ups and to preserve the conversational state rather than returning a single isolated answer.

Add alternative phrasings is focused on increasing coverage for how users might ask the same single question and it does not create the branching prompts or state tracking needed for multi turn conversations.

Enable chit chat would add casual small talk capabilities and scripted conversational replies but it is not the mechanism used to build structured multi turn flows within Azure Language question answering.

Enable active learning helps identify ambiguous or low confidence matches for human review so the knowledge base can improve over time and it does not enable the runtime behavior of maintaining context across multiple dialogue turns.

Look for keywords like multi turn or context in the question and pick the option that explicitly creates follow up paths or prompts to continue the conversation.