Microsoft Azure AI Engineer Practice Exams

The container sends usage metrics to the Azure AI Services resource for billing and client applications do not need to supply a subscription key while request data is processed locally.

This option is correct because Azure AI Services containers process request payloads on the local Docker host and do not automatically forward full request content to the Azure service for processing or storage. The container still reports usage metrics and telemetry to the Azure AI Services resource so that billing and usage tracking can occur, and the container can manage the link to the Azure resource so client applications do not need to include a subscription key for each request.

All request payloads sent to the container are automatically forwarded to the Azure AI Services endpoint for processing and storage is incorrect because containers are designed to run model inference locally and they do not by default transmit entire request payloads to the cloud for processing or long term storage.

The container runs completely offline and no usage metrics are ever reported to Azure is incorrect because containers do report usage metrics and minimal telemetry to the Azure AI Services resource for billing and monitoring when the container is configured to associate with an Azure resource.

Client applications must include a subscription key and authenticate with the Azure resource before the container can be used is incorrect because client apps typically send requests to the local container endpoint and the container itself handles the association with the Azure resource for billing and telemetry, so clients do not need to supply the Azure subscription key for each request.

When answering container questions remember that containers run inference locally and do not send full request payloads by default. Look for mention of usage metrics or telemetry to determine if the cloud service is still involved for billing.

The correct answer is JSON manifest in the project storage account container.

The JSON manifest in the project storage account container contains the labels and the asset references that represent the multilabel annotations for the project and it is saved alongside other training artifacts so the service can use it to reproduce or export the model.

The manifest format supports multiple labels per example and nested metadata which is why a structured JSON is used rather than a flat file. The project storage account container is the managed location for those artifacts which makes it the canonical place to find the training data for a completed model.

CSV file in project Blob Storage container is incorrect because CSV files do not capture multilabel annotations and related metadata as reliably and the service stores a structured manifest instead.

Files in an Azure Data Lake storage account is incorrect because the training manifest is placed in the project storage account container by the service unless you explicitly export the project to another storage target and ADLS is not the default location.

When you need to identify where a managed Azure service keeps its training assets look for a service specific project container and remember that multilabel annotations are typically stored in a JSON manifest rather than a flat CSV.

The correct answer is Azure AI Document Intelligence.

Azure AI Document Intelligence is designed to extract both printed and handwritten text and to capture key value pairs from documents such as invoices. The service includes prebuilt invoice and form models and layout analysis that return structured fields and confidence scores which makes it suitable for a vendor portal that accepts scanned invoice images.

Google Document AI is a capable document parsing product but it is a Google Cloud service and not the Azure solution Jordan was asked to pick for an Azure deployment.

Azure Cognitive Search is focused on indexing and searching content and on adding search experiences to applications and it does not provide the specialized prebuilt invoice key value extraction that Document Intelligence offers.

Azure Read Assistant is not the documented Azure service for invoice level OCR and key value extraction and it does not map to the prebuilt invoice models available in Document Intelligence.

None of the listed options are appropriate is incorrect because an appropriate Azure service is listed and it is Azure AI Document Intelligence.

When the question asks about extracting text from scanned documents and capturing key value pairs look for services that mention document or form intelligence and prebuilt invoice models. Also confirm the service belongs to the same cloud provider named in the question.

The correct option is azure-ai-language-conversations.

The azure-ai-language-conversations package is the Azure SDK client library that targets the Language service conversations and conversation analysis features and it is intended for calling intent classification models. This package provides the client types and methods to submit user utterances and receive predicted intents and extracted entities from conversation or intent classification models.

azure-ai-textanalytics is focused on general text analytics tasks such as sentiment analysis, key phrase extraction, and named entity recognition and it does not provide the conversations intent classification APIs that the Language service conversations functionality exposes.

azure-cognitiveservices-speech is the Speech SDK for speech to text, text to speech, and speech translation and it is not used to call intent classification or conversation analysis models in the Language service.

When a question asks about intent classification or conversation analysis look for package names that mention language or conversations since those packages target the Language service conversation features.

Inclusiveness and accessibility and Fairness and non discrimination are correct.

Fairness and non discrimination focuses on detecting and mitigating bias so that analytics and model outputs do not produce systematically worse outcomes for particular regions or demographic groups. Monitoring fairness means disaggregating performance and outcome metrics by relevant attributes and applying corrective measures when disparities arise.

Inclusiveness and accessibility is about ensuring the system works for diverse user needs and contexts so that image analysis and chat processing do not exclude or misrepresent populations. Monitoring for inclusiveness includes validating coverage across languages, regions, devices, and accessibility needs and ensuring that models perform acceptably for all covered groups.

Transparency and explainability is valuable for understanding model behavior and for communicating decisions to stakeholders but it does not by itself ensure equitable outcomes across groups.

Accountability and governance provides oversight and processes to respond to harms and to enforce policies but it is not the principle that directly guides monitoring for disparate impacts across demographic groups.

Reliability and safety addresses robustness and the prevention of failures or harmful outputs and it complements equity work but it does not specifically target monitoring for fairness or inclusive coverage.

When a question asks about monitoring for equitable outcomes focus on principles that directly address disparate impacts and user coverage and look for answers that mention fairness or inclusion.

Upload all images and request suggested tags then review and accept them is correct.

This option leverages Custom Vision to automatically suggest tags for large numbers of unlabeled images so you can accept or correct predictions and then retrain the model quickly. Using the suggested tags feature lets you handle a bulk upload and then perform a focused review rather than spending hours labeling every photo by hand.

Upload images organized by category and tag each photo manually then review suggested tags is not ideal because manually tagging 1200 photos is time consuming and defeats the purpose of using the service to accelerate labeling.

Prelabel images with an external script and import the labeled dataset is also less efficient in this scenario because writing and validating a labeling script adds overhead and you still need to verify labels. Using the built in suggestion and review workflow is faster for large unlabeled datasets.

For large unlabeled image sets pick the option that uses the service’s suggested tags or auto tagging and always review the suggested labels before retraining the model.

The correct option is Text-independent verification.

This option allows verification of a speaker using natural or spontaneous speech and it does not require scripted phrases or fixed text during enrollment or verification. That behavior makes it suitable for meeting recordings where participants speak freely and you need to confirm a claimed identity from the audio.

Speaker diarization is incorrect because diarization only segments audio and groups speech by speaker turns and it does not authenticate or verify a speaker identity. Diarization helps you know who spoke when but it cannot confirm that a voice belongs to a specific enrolled person.

Text-dependent verification is incorrect because it requires the same passphrase or fixed text during enrollment and verification. That dependency makes it unsuitable for unscripted meeting analysis where participants do not speak predetermined phrases.

Speaker identification is incorrect because identification attempts to determine which enrolled speaker from a set produced an audio sample and it is a one to many process. Identification is not the right fit when the goal is to verify a claimed identity in recordings without scripted speech.

When a scenario asks about voice profiles check whether the task is one to one verification or one to many identification and check whether enrollment requires fixed text. Prioritize text-independent verification when recordings contain spontaneous speech.

The correct option is Fine tuning. This approach adapts an Azure OpenAI base model by training it on labeled image captions so the model learns domain specific language patterns and produces outputs that better match the captioning style and vocabulary you provide.

Embeddings are vector representations used for similarity search and retrieval. They help with finding relevant captions or images but they do not modify model weights and therefore do not adapt the base model through training on labeled examples.

Prompt engineering involves crafting inputs to get better responses from the existing model. It can improve results at inference time but it does not train the model on labeled image captions and so it cannot permanently incorporate domain specific knowledge in the way a training based adaptation can.

When a question says the model is being “trained” on labeled examples choose Fine tuning rather than options that only change inputs or use vector representations.

text-embedding-ada-002 is correct.

text-embedding-ada-002 is an embeddings model that converts text into numeric vectors that capture semantic meaning. You can embed each of the 15,000 files and then compute cosine similarity between the query embedding and the file embeddings to find which files contain the target phrases. Embedding models are the right choice when you need vector representations for semantic search and similarity tasks.

Be aware that embedding model names have evolved and newer embedding models may appear on future exams, but the key idea is to choose an embeddings model rather than a chat or completion model.

GPT-4-32k is a large generative model that is designed for text generation and long-context reasoning. It does not provide the efficient embedding vectors you need for cosine similarity based vector search and it would be unnecessarily expensive and complex for this task.

GPT-3.5 Turbo is optimized for chat and completions and not for producing dense vector embeddings for semantic similarity. It is not the appropriate choice when you must perform cosine similarity on embeddings.

GPT-4 is also a generative model intended for high-quality text generation and reasoning. It is not the correct model for embedding generation and vector similarity searches.

When a question asks about cosine similarity or vector search pick an embeddings model rather than a chat or completion model.

The correct options are text, language, and id.

The text field is required because it contains the raw content that the entity extraction model analyzes and without it there is nothing to process.

The language field is required because it informs the endpoint which language rules and tokenization to apply so the model can correctly recognize entities.

The id field is required because it gives each JSON entry a stable identifier so results can be correlated back to the original input in batch jobs or downstream systems.

The timestamp field is not required for entity extraction because time metadata is optional and it is not needed for identifying or extracting entities. It may be present for auditing or ordering but it is not a mandatory input to the extraction endpoint.

When you read field lists focus on what the model needs to perform extraction and what is only metadata. Pay attention to words like required and look for fields that carry content rather than optional context.

All exam questions are from certificationexams.pro and my Azure AI Udemy course.

The correct sequence is 2 then 1 then 3.

The first call in that sequence retrieves the trained iteration metadata and identifier from the 2 then 1 then 3 source project on cvstaging so you know which iteration to move. The next call exports or downloads the trained object detection model artifact from the staging resource. The final call uploads or creates the iteration in the cvlive resource so the model is available in production.

2 then 3 then 1 is incorrect because it attempts to perform the import or finalization step before the export or download step and therefore would not have the model artifact to import.

2 then 1 then 5 is incorrect because the final call in that sequence is not the required import or create action on the production resource and so it does not complete the move into cvlive.

2 then 1 then 4 is incorrect because the final call in that option does not perform the proper import or creation of the iteration in cvlive and so it fails to place the trained model into the production resource.

On exam questions about moving models map each API call to a stage. Think in terms of retrieve iteration metadata, then export or download, then import or create in the target resource and choose the sequence that matches those three stages.

No the usage attribute is on the top level response object is correct.

The Azure OpenAI chat completions response exposes usage as a top level metadata object that summarizes token counts and related billing information. The messages or choices in the response contain role and content fields and they do not carry the usage counts. Placing usage at the top level makes it easy to read total token usage without inspecting each message.

Yes the usage is nested inside the message content is incorrect because usage is not embedded inside any message or choice body. That option would imply the API mixes metadata into the message payload which is not how the chat completions response is defined.

When the exam asks where data appears in a response look for fields that represent metadata like usage at the top level and fields that represent content like messages under choices.

The correct options are Configure AIHubEast network access to restrict traffic to AppNet2 and a specific subnet and Enable a service endpoint for Microsoft.CognitiveServices on AppNet2.

Applying Configure AIHubEast network access to restrict traffic to AppNet2 and a specific subnet ensures the Azure AI service accepts connections only from the specified virtual network and subnet. This blocks public access at the service level so only internal resources in AppNet2 can reach AIHubEast.

Enabling Enable a service endpoint for Microsoft.CognitiveServices on AppNet2 extends the virtual network identity to the Cognitive Services endpoint over the Azure backbone. This keeps traffic on Microsofts network and is straightforward to configure at the subnet level so it reduces administrative overhead compared with more complex alternatives.

Together these two settings restrict access to the desired internal subnet and block public access while keeping configuration and ongoing management simple.

Enable a private endpoint for AIHubEast is not chosen because a private endpoint introduces private IPs, DNS updates, and extra lifecycle management. That approach gives stronger isolation but it increases operational complexity which contradicts the requirement to keep administrative overhead low.

Create an additional subnet inside AppNet2 is unnecessary because AIHubEast can be restricted to an existing specific subnet in AppNet2, and adding a subnet does not directly block public access by itself.

Modify the Access control IAM settings on AIHubEast is not relevant for blocking network access because IAM controls authorization and not the network path. Changing IAM does not stop public network traffic to the service.

When a question asks to block public access and minimize operational work prefer network restrictions and service endpoints for supported PaaS services and reserve private endpoints for cases that require stronger isolation despite extra DNS and management.

Add training utterances for languages that show low accuracy in the model is the correct choice.

Add training utterances for languages that show low accuracy in the model improves multilingual accuracy because the model learns the actual phrasing and patterns used by speakers of that language. Adding examples is usually the smallest development effort since you extend the existing project and retrain rather than building and maintaining new infrastructure.

Create a separate CLU project for each language is not ideal because it multiplies maintenance work and requires duplicating training, testing, and deployment for each language. That approach increases overhead and is not minimal effort.

Use Azure Translator to translate incoming queries into the model training language before sending them to CLU can introduce translation errors and extra latency which can reduce intent accuracy. It also requires building and operating a translation pipeline which adds development and operational effort.

Use a language detection step to route queries to language specific CLU projects is also suboptimal because it creates more projects to manage and adds routing complexity. Language detection plus per‑language projects increases engineering work compared with extending a single multilingual model with extra utterances.

When you see low accuracy in a particular language, first add representative user utterances in that language and retrain. This usually yields the best improvement with the least development overhead.

The correct answer is table projection.

table projection tells the Azure AI Search indexer to emit extracted table data in a rows and columns format so analytics tools like Power BI can consume it with minimal transformation. This projection converts document or AI extracted tables into tabular output that is easier to map into a Power BI dataset and it reduces the need to write custom extraction or reshaping code.

object projection is incorrect because it is used to project nested JSON objects into the index schema and not to produce a flattened tabular output suitable for direct consumption by Power BI.

projection group is incorrect because grouping projections affects how multiple projections are applied together and does not by itself create the row and column structure that Power BI needs.

field mapping is incorrect because mappings only map source fields to index fields and they do not transform or render extracted table data into a tabular set of rows and columns.

When a question asks for output in a tabular form for analytics think about features that explicitly produce rows and columns. Remember that table projection is the Azure AI Search feature that creates tabular outputs for tools like Power BI.

The correct answer is Only when he needs to match the exact words the user spoke.

Pattern matching is appropriate when you must detect text verbatim and you cannot tolerate the model interpreting synonyms or paraphrases. In that scenario a deterministic pattern or exact string match gives precise results and avoids the variability that a conversational language model introduces.

When he requires an entity extracted by a machine trained model is incorrect because that statement describes the opposite situation. If he needs machine trained extraction then a conversational language model or ML based entity extractor is the right choice.

Regex pattern matching is incorrect as an answer because it names a technique rather than the decision criterion the question asks for. Regex can be a tool to implement pattern matching but the question asks when to choose pattern matching instead of a conversational model.

Azure Conversational Language Understanding is incorrect because it refers to a conversational machine learning service that generalizes from examples. That service is what you would use when you want the model to infer intent or extract entities from varied phrasing rather than match exact words.

Look for keywords like exact or verbatim in the question to decide that deterministic pattern matching is required instead of a conversational model.

The correct answer is Yes.

When you construct a SpeechConfig with your subscription key and region and create an AudioConfig by calling AudioConfig.FromWavFileOutput with the path “records/audio_out_042.wav” you are instructing the Speech SDK to write synthesized audio to that WAV file. Instantiating a SpeechSynthesizer with those configs and calling SpeakTextAsync causes the service to synthesize the text and save the resulting audio to the specified file path.

No is incorrect because the code path described explicitly uses AudioConfig.FromWavFileOutput which directs output to a WAV file. The only reasons a file would not appear are environmental issues such as an invalid path or lack of write permissions rather than the SDK behavior itself.

On the exam look for API names like AudioConfig.FromWavFileOutput as they usually indicate whether audio is written to a file or played to speakers and also check that the question assumes the application has proper write permissions to the target path.

Object detection is correct because it returns the locations of objects in an image and so it can detect defects and count items in packaging.

Object detection uses bounding boxes to localize each instance of an object and it can return multiple detections per image so you can both identify defects and tally items. This localization capability is what makes Object detection suitable for tasks that require coordinates and counts rather than just labels.

Image classification is incorrect because it assigns labels to an entire image or to the predominant content and it does not provide coordinates for object locations. It cannot tell you where a defect is or how many items are present.

Image segmentation is incorrect because it produces pixel level masks that outline regions rather than returning bounding box coordinates for individual objects. Segmentation gives precise shapes but it is not the typical choice when you need simple object counts with boxes.

Compact classification is incorrect because it refers to an optimized, lightweight form of classification model for edge deployment and it still performs classification only. It does not provide localization data or per object coordinates that are needed to detect defects and count items.

When a question mentions locations or counts look for options that provide localization. In Azure Custom Vision that usually means Object detection rather than classification.

The correct option is Resource 1 corresponds to characteristics 2 and 3.

A multi service Azure AI resource is designed to provide consolidated billing for the services that are used and to present a single key and endpoint that grants access to multiple Azure AI capabilities. Those behaviors match Resource 1 corresponds to characteristics 2 and 3 and explain why the multi service resource is mapped to those characteristics.

Resource 2 corresponds to characteristics 2 and 3 is incorrect because consolidated billing and a single key and endpoint describe a multi service resource rather than a single service resource. Resource 2 is a single service resource so it would not provide those multi service conveniences.

Resource 1 corresponds to characteristics 1 and 4 is incorrect because characteristic 1 describes a single service access model and that does not apply to a multi service resource. Characteristic 4 is about an available free tier and that is not a defining property of the multi service resource in this mapping.

Resource 2 corresponds to characteristics 1 and 4 is incorrect as an answer in this question because the expected mapping for the multi service and single service examples focuses on the credential and billing models. While characteristic 1 fits a single service resource, characteristic 4 is not a guaranteed characteristic for the single service resource in the context of this exam style question.

Resource 1 corresponds to characteristics 2 and 4 is incorrect because characteristic 4 is not a defining trait of the multi service resource and the mapping omits the single key and endpoint behavior captured by characteristic 3.

Resource 1 corresponds to characteristics 1 and 2 is incorrect because characteristic 1 denotes a single service credential model which contradicts the multi service nature of Resource 1, and the pairing misses the single key and endpoint aspect that is central to the multi service resource.

Look for language about a single key and endpoint to identify multi service Azure AI resources and look for service specific keys to identify single service resources.

reportA.docx scanB.pdf photoC.jpeg graphicD.png is correct. This option lists a DOCX file, a PDF file, a JPEG image and a PNG image and those common document and image formats are supported by Azure AI Content Understanding for analysis.

Azure AI Content Understanding accepts both document formats and image formats so it can extract text and structure from Word documents and from scanned images or photos. The list in the correct option matches the supported DOCX, PDF, JPEG and PNG types that the service analyzes.

scanB.pdf photoC.jpeg graphicD.png is incorrect because it leaves out the DOCX document format which is included in the supported types and is present in the correct option.

reportA.docx scanB.pdf photoC.jpeg graphicD.png iconE.webp is incorrect because it adds a .webp file which is not part of the standard supported input formats for Content Understanding and so that option lists a file type the service does not accept.

When answering format support questions focus on the file extensions listed and compare them to the service documentation for supported file types rather than being distracted by extra file names.

Cameron McKenzie is an AWS Certified AI Practitioner, Machine Learning Engineer, Solutions Architect and author of many popular books in the software development and Cloud Computing space. His growing YouTube channel training devs in Java, Spring, AI and ML has well over 30,000 subscribers.