Free Practice Questions for the Microsoft Azure AI Engineer Associate AI-102 Exam (2026 Updated)

In Custom Question Answering (part of Azure AI Language), if you want to define alternative words or spellings for a specific term (e.g., testing , trials , evaluate ), you must use the alterations property.

Each alteration object contains a list of alternative values for the same meaning.

Example:

{

" alterations " : [

{

" alterations " : [

" testing " ,

" trials " ,

" evaluate "

]

}

]

}

phrases is used for pattern matching in question-answer pairs.

synonyms is used for synonyms within LUIS (Language Understanding), not for QnA alterations.

value is used inside synonyms definitions in some contexts but not here.

Correct JSON Payload Completion:

The JSON should be:

{

" alterations " : [

{

" alterations " : [

" testing " ,

" trials " ,

" evaluate "

]

}

]

}

Verified Answer:

Left box: " alterations "

Right box: " alterations "

Microsoft References:

Manage alterations in Custom Question Answering (Azure AI Language)

Authoring API reference for alterations

Cognitive Insights Widget URL

https: / /www.videoindexer.ai/embed/insights/ {accountId}/{videoId}/ ?w idgets=people,keywords & controls=search

Player Widget URL

https: / /www.videoindexer.ai/embed/player/ {accountId}/{videoId}/ ?s howcaptions= true & captions=en- US

Comprehensive Detailed Explanation

The requirements are:

Ensure that users can search for keywords.

In the Cognitive Insights Widget, the controls parameter can be set to search to allow keyword search.

Display the names and faces of people in the video.

The widgets parameter in the Cognitive Insights Widget determines which insights are shown.

To show people and keywords, use people,keywords .

Show captions in the video in English (United States).

For the Player Widget, captions are controlled by two parameters:

showcaptions=true enables captions.

captions=en-US specifies the language.

Final Placement of Values

Cognitive Insights Widget

widgets=people,keywords

controls=search

Player Widget

showcaptions=true

captions=en-US

Microsoft References

Embed Video Indexer widgets

Video Indexer parameters and customization

“Protected material detection” in Content Safety is designed to flag model outputs that match copyrighted text/code (e.g., lyrics, articles, recipes, GitHub code). It is not used to tune or test content filter configurations for safety categories like hate/sexual/violence/self-harm. Therefore, it does not meet the requirement to optimize content filters by running tests on sample questions.

For the category field in Azure Cognitive Search:

searchable → allows users to search by words in the category field.

facetable → enables drill-down filtering (facets).

retrievable → ensures the field appears in search results.

filterable/sortable/key are not required here based on the scenario.

Correct Answer: A

The exhibit shows incremental increases by quarter leading to a cumulative running total , which is best represented with a waterfall chart .

Ribbon chart shows ranking changes over time, not cumulative totals.

Bar chart shows simple comparisons, not running totals.

Decomposition tree is for drill-down analysis, not running totals.

Correct Answer: A

Language Understanding: An AI service that allows users to interact with your applications, bots, and IoT devices by using natural language.

QnA Maker is a cloud-based Natural Language Processing (NLP) service that allows you to create a natural conversational layer over your data. It is used to find the most appropriate answer for any input from your custom knowledge base (KB) of information.

Text Analytics: Mine insights in unstructured text using natural language processing (NLP)—no machine learning expertise required. Gain a deeper understanding of customer opinions with sentiment analysis. The Language Detection feature of the Azure Text Analytics REST API evaluates text input

Comprehensive Detailed Explanation along with All References available from Microsoft Azure AI Solution at the end of the explanation

Typing: In a .chat transcript, you can simulate the bot’s typing indicator with the Typing activity, optionally followed by a [Delay=...] . This matches the intent to pause before sending the next message.

AttachmentLayout=carousel: When sending multiple attachments (the two watch images), set the layout to carousel so the Emulator displays the attachments horizontally as a card carousel.

Adaptive Card attachment: The JSON file cards\watchProfileCard.json represents an Adaptive Card; specify adaptivecard so the Emulator renders the card.

Resulting .chat snippet (essentials):

user =User1

bot=watchbot

user : I want a new watch.

bot: [Typing][Delay= 3000 ]

bot: I can help you with that! Let me see what I can find.

bot: Her e ' s what I found.

bot:

[AttachmentLayout=carousel]

[Attachment=https://contoso.blob.core.windows.net/watch01.jpg]

[Attachment=https://contoso.blob.core.windows.net/watch02.jpg]

user: I like the first one.

bot: Sure, pulling up more information.

bot: [Attachment=cards\watchProfileCard.json adaptivecard]

user: That ' s nice! Thank you.

bot: Sure, you are most welcome!

References

Bot Framework Emulator transcript ( .chat ) activities and typing indicator: Bot Framework Activity schema; Emulator supports typing and delay in transcripts.

https://learn.microsoft.com/azure/bot-service/bot-service-debug-emulator

https://learn.microsoft.com/azure/bot-service/rest-api/bot-framework-rest-connector-api-reference#activity-object

Attachment layout ( carousel / list ) for multiple attachments:

https://learn.microsoft.com/azure/bot-service/nodejs/bot-builder-nodejs-send-rich-cards#carousel-layout

Adaptive Cards in Bot Framework:

https://learn.microsoft.com/adaptive-cards/get-started/bots

https://learn.microsoft.com/azure/bot-service/bot-builder-howto-add-media-attachments#adaptive-cards

To restrict access so that only resources in vnet1 can access the Language service:

You configure the virtual network settings on the Language service resource.

This enables private endpoint integration with the VNet.

NSGs and Azure Firewall apply to subnets/resources but do not configure the service resource itself.

Containers can run Language service locally but are not needed here.

Correct Answer: C

You are deploying a Language Understanding (LUIS) container. When running LUIS in Docker, the model must first be exported from the Language Understanding portal, and then provided to the container at runtime.

Step-by-step reasoning:

From the Language Understanding portal, export the solution as a package file.

The trained LUIS model must be exported from the portal into a .json package file.

This is required because the container cannot access the hosted service directly.

From the host computer, move the package file to the Docker input directory.

Containers expect the model to be available locally.

The package file is placed into the input directory that the container maps for models.

From the host computer, run the container and specify the input directory.

When starting the container, you specify --volume < local_input > :/input so the container has access to the package file.

This makes the exported model available inside the container for processing.

Why not the other options?

Retain the model in the portal is not sufficient; the container cannot pull directly from the cloud.

Build the container and specify the output directory is not required; the container image is already available from MCR and is not custom-built for this step.

Correct Answer Order:

Export the solution as a package file.

Move the package file to the Docker input directory.

Run the container and specify the input directory.

Microsoft References

Run LUIS containers

Use containers with Azure AI services

To maximize answer accuracy (reduce randomness and hallucinations) with Azure OpenAI GPT-3.5, you should lower the sampling randomness. The primary control for this is temperature :

Temperature controls how “creative” or random the model is.

Lower values → more deterministic, focused, and conservative outputs.

temperature = 0 makes the model pick the highest-probability tokens, which typically yields the most accurate and repeatable responses for transactional bots like booking/reservations.

Top P (nucleus sampling) is an alternative to temperature. Microsoft recommends using one or the other. While reducing Top P also narrows randomness, the clearest and most common approach for maximizing accuracy is setting temperature to 0. Setting Top P to 0 is generally not recommended and may be invalid depending on the SDK.

System message tuning (Option D) helps with tone, role, and policy adherence, but does not itself guarantee factual accuracy if decoding randomness remains high.

Including the travel agent’s database (Option A) would require Azure OpenAI on your data or a RAG pattern. That can improve grounding of facts, but it’s not a model parameter change and is outside the scope of simply configuring GPT-3.5; moreover, accuracy still benefits from low temperature even when using your own data.

Therefore, the best single action here is: set Temperature to 0.

Microsoft Azure AI Solution References

Azure OpenAI Service – Chat Completions parameters (Temperature and Top P): “Lower values like 0.2 will make the output more focused and deterministic… use either temperature or top_p but generally not both.”

https://learn.microsoft.com/azure/ai-services/openai/reference#chat-completions

Prompt engineering with Azure OpenAI – controlling randomness and determinism:

https://learn.microsoft.com/azure/ai-services/openai/concepts/prompt-engineering

Azure OpenAI on your data (grounding with enterprise data/RAG):

https://learn.microsoft.com/azure/ai-services/openai/concepts/use-your-data

=========================

Comprehensive Detailed Explanation

The requirement:

You have an Azure AI Document Intelligence resource (formerly Form Recognizer).

You want the app to detect handwritten content inside PDF files.

Step 1 – Choosing the correct model

prebuilt-document: Extracts structure and text from documents, but not optimized specifically for handwriting.

prebuilt-contract: Designed for contracts, not general handwriting.

prebuilt-read: Extracts text (printed and handwritten) from documents. Supports handwritten content recognition.

Therefore, to analyze handwritten content, the correct choice is prebuilt-read.

Step 2 – Confidence threshold

The style.Confidence value indicates the confidence that the text is handwritten.

A typical threshold used is 0.75 (75%).

This ensures that only results with reasonably high confidence are considered handwritten.

Values like 0.1 would be too low (allowing false positives), while 1.0 might miss valid handwriting because it requires absolute certainty.

Correct Answer:

Model: prebuilt-read

Confidence threshold: 0.75

Comprehensive Detailed Explanation

You’re adding multiple images for each person into a PersonGroup using the Face API SDK.

File.OpenRead(imagePath) returns a Stream, so the using declaration must declare Stream t = File.OpenRead(imagePath) .

To upload an image from a stream to a person in a person group, call

faceClient.PersonGroupPerson.AddFaceFromStreamAsync(personGroupId, personId, t);

Putting it together:

Parallel.For( 0 , PersonCount, async i = >

{

Guid personId = persons[i].PersonId;

string personImageDir = $ " /path/to/person/ {i} /images " ;

foreach ( string imagePath in Directory.GetFiles(personImageDir, " *.jpg " ))

{

using (Stream t = File.OpenRead(imagePath))

{

await faceClient.PersonGroupPerson.AddFaceFromStreamAsync(

personGroupId, personId, t);

}

}

});

This loops the images for each person, opens each as a stream, and adds the face to the specified person in the person group.

References

Microsoft Docs – Face API: Add a face to a person ( AddFaceFromStreamAsync , AddFaceFromUrlAsync ) in PersonGroupPerson operations.

https://learn.microsoft.com/azure/ai-services/face/ (see SDK/PersonGroupPerson methods)

NET File.OpenRead returns Stream .

https://learn.microsoft.com/dotnet/api/system.io.file.openread

In Azure Cognitive Search, a knowledge store is an optional capability that allows you to persist the enriched data from a skillset (enrichment pipeline) into an Azure Storage account. This is often used when you want to further analyze or visualize enriched content, e.g., sentiment analysis results from social media posts.

When defining a knowledge store, you need:

storageConnectionString

This is required so the enrichment pipeline can connect to your Azure Storage account.

Without it, the knowledge store cannot persist any data.

storageContainer

This specifies the container in Blob Storage where enriched documents or projections will be stored.

This allows you to persist the original documents and enriched outputs such as sentiment analysis scores.

Other options (not correct in this case):

tables → This is a projection type used within a knowledge store to structure output in Table Storage, but it’s not required in the definition itself.

files → Another projection type for storing enriched documents as files in Blob Storage, but again, not a required field in the definition itself.

objects → A projection type used for representing enriched documents as JSON objects in Blob Storage.

Thus, while tables, files, and objects are used inside projections within a knowledge store, the definition itself always requires a storage connection string and storage container.

Correct Answer:

A. storageContainer

C. storageConnectionString

Microsoft References

Knowledge store in Azure Cognitive Search

How to define a knowledge store

Correct Sequence:

Add the prebuilt domain ToDo

Add a new application

Train the application

Publish the application

Comprehensive Detailed Explanation

You are building a Conversational Language Understanding (CLU) model for a chatbot that supports task tracking. To minimize development time, the best approach is to use a prebuilt domain (like ToDo) instead of building intents/entities manually.

Steps in correct order:

Add the prebuilt domain ToDo

Prebuilt domains in CLU provide ready-to-use intents and entities for common scenarios (e.g., calendar, email, ToDo).

This significantly reduces development time since you don’t need to create all intents/entities manually.

Add a new application

You must first create a CLU application that uses the prebuilt domain.

The application acts as the container for intents, entities, and utterances.

Train the application

Training is required after adding intents, entities, and utterances (or using a prebuilt domain).

This ensures the model can classify and extract information correctly.

Publish the application

Once trained, you must publish the application so it can be accessed via an endpoint and integrated into the chatbot.

Incorrect Actions in this Case

Add example utterances → Not required if you’re using the prebuilt ToDo domain (utterances are already included).

Correct Sequence:

Add the prebuilt domain ToDo

Add a new application

Train the application

Publish the application

Microsoft References

Conversational Language Understanding (CLU) overview

Prebuilt domains in CLU (LUIS ToDo example)

Train and publish CLU models

Scenario Recap:

Application: Flower classification.

Solution: Create a new model, and then upload the new images and labels.

Analysis:

Creating a new model from scratch is not the correct approach for adding species to an existing classifier .

The goal is to extend the existing model, not reset and rebuild a new one.

This would waste training data and is not aligned with Custom Vision usage.

Correct Answer: B. No

Comprehensive Detailed Explanation

You have an Azure Cognitive Search service. Query volume has increased, and some search requests are now being throttled. This means your current tier is no longer sufficient to handle the query traffic.

Why does migrating to a higher tier work?

Each Azure Cognitive Search pricing tier (Basic, Standard S1/S2/S3, Storage Optimized, etc.) provides different resource limits for queries per second (QPS), indexing throughput, and storage.

Throttling occurs when query traffic exceeds the capacity limits of the current tier.

Moving to a higher tier increases:

The allowed query units (QUs),

Maximum queries per second (QPS),

and available compute resources.

Therefore, migrating to a higher tier reduces the likelihood of throttling and supports the increased query volume.

Why not other solutions?

Simply adding replicas can help scale out query workloads, but the question specifically asks whether moving to a higher tier meets the goal—and it does.

Using indexer scaling or adjusting query patterns might help, but they are not direct answers to the throttling caused by insufficient service tier capacity.

Correct Answer: A. Yes

Microsoft References

Azure Cognitive Search service limits by tier

Scale resources in Azure Cognitive Search

Azure Cognitive Search pricing tiers

By default, Azure AI Language may log request and response data for monitoring.

Setting the loggingOptOut parameter to true ensures that input data is not persisted after analysis.

piiCategories specifies which PII types to detect.

showStats enables request/response statistics.

model-version specifies which model version to use.

Correct Answer: A

You’re indexing content with Azure AI Search and you want to recognize and retrieve properties from invoices (such as invoice number, vendor, dates, totals, line items). The Azure service designed for structured information extraction from forms and documents is Azure AI Document Intelligence (formerly Form Recognizer).

In an Azure AI Search enrichment pipeline, you add a skillset. To pull structured fields from invoices, include a skill that invokes Document Intelligence’s prebuilt Invoice model. This model already extracts common invoice fields, so you don’t have to train anything—meeting the goal of minimal effort. You can wire this in using the built-in Document Intelligence skill for Search, or as a custom Web API skill that calls the Document Intelligence REST API, and then map the extracted fields into your index for retrieval and filtering.

Why the others are not suitable:

Azure OpenAI is for generative/LLM tasks, not schema-level invoice field extraction.

Azure AI Immersive Reader improves reading accessibility; it doesn’t extract business fields.

Azure Custom Vision is for image classification/object detection, not form/key-value extraction.

Microsoft References (Azure AI Solution)

Azure AI Search skillsets – add AI enrichment and custom skills (Web API skill).

Document Intelligence skill for Azure AI Search (formerly Form Recognizer skill) – integrate extraction into indexing.

Prebuilt Invoice model – fields extracted (vendor, invoice ID/date, totals, taxes, line items, etc.).

Azure AI Document Intelligence overview – key-value extraction and table/line-item parsing from documents.

Scenario: When the response confidence score is low, ensure that the chatbot can provide other response options to the customers.

When no good match is found by the ranker, the confidence score of 0.0 or " None " is returned and the default response is " No good match found in the KB " . You can override this default response in the bot or application code calling the endpoint. Alternately, you can also set the override response in Azure and this changes the default for all knowledge bases deployed in a particular QnA Maker service.

Choosing Ranker type: By default, QnA Maker searches through questions and answers. If you want to search through questions only, to generate an answer, use the RankerType=QuestionOnly in the POST body of the GenerateAnswer request.

Can search content in different formats, including video

Audio and video insights (multi-channels). When indexing by one channel, partial result for those models will be available.

Keywords extraction: Extracts keywords from speech and visual text.

Named entities extraction: Extracts brands, locations, and people from speech and visual text via natural language processing (NLP).

Topic inference: Makes inference of main topics from transcripts. The 2nd-level IPTC taxonomy is included.

Artifacts: Extracts rich set of " next level of details " artifacts for each of the models.

Sentiment analysis: Identifies positive, negative, and neutral sentiments from speech and visual text.

Step 1: Create a new project and load sample documents

Create a new project. Projects store your configurations and settings.

Step 2: Label the sample documents

When you create or open a project, the main tag editor window opens.

Step 3: Train a custom model.

Finally, train a custom model.

The requirements state:

All AI solution responses must have a confidence score ≥ 70%.

If the response confidence score is < 70%, the response must be improved by human input.

Members of the Consultant-Bookkeeper group must be able to process financial documents, which includes performing manual review when the AI confidence is below threshold.

Looking at the provided JSON:

" fields " : {

" ReceiptType " : {

" type " : " string " ,

" valueString " : " Itemized " ,

" confidence " : 0.672

},

" MerchantName " : {

" type " : " string " ,

" valueString " : " Tailwind " ,

" confidence " : 0.913

}

}

ReceiptType.confidence = 0.672 → this is below 0.7.

MerchantName.confidence = 0.913 → this is above 0.7.

Therefore, the correct condition for triggering manual review is:

documentResults.fields. < field > .confidence < 0.7

Now, evaluating the options:

A. documentResults.docType == " prebuilt:receipt "

Always true for receipts, does not check confidence. Not correct.

B. documentResults.fields.*.confidence < 0.7

This is the correct general expression: trigger manual review whenever any field confidence is below 0.7.

C. documentResults.fields.ReceiptType.confidence > 0.7

This would bypass manual review when ReceiptType has high confidence. The requirement is to trigger review when confidence < 0.7, so this is the opposite.

D. documentResults.fields.MerchantName.confidence < 0.7

This only checks one field (MerchantName). In the JSON, MerchantName has confidence 0.913 ( > 0.7), so this condition would not trigger, but ReceiptType clearly needs review. Too narrow.

Correct Answer: B. documentResults.fields.*.confidence < 0.7

Microsoft References

Azure AI Document Intelligence – Confidence scores

Human-in-the-loop for Document Intelligence

Box 1: " categories " : [ " Locations " , " Persons " , " Organizations " ],

Locations, Persons, Organizations are in the outputs.

Scenario: Contoso plans to develop a searchable knowledgebase of all the intellectual property

Note: The categories parameter is an array of categories that should be extracted. Possible category types: " Person " , " Location " , " Organization " , " Quantity " , " Datetime " , " URL " , " Email " . If no category is provided, all types are returned.

Box 2: { " name " : " entities " }

The include wikis, so should include entities in the outputs.

Note: entities is an array of complex types that contains rich information about the entities extracted from text, with the following fields

name (the actual entity name. This represents a " normalized " form)

wikipediaId

wikipediaLanguage

wikipediaUrl (a link to Wikipedia page for the entity)

etc.

Box 1: Cognitive Service User

Ensure that the members of a group named Management-Accountants can approve the FAQs.

Approve=publish.

Cognitive Service User (read/write/publish): API permissions: All access to Cognitive Services resource except for ability to:

1. Add new members to roles.

2. Create new resources.

Box 2: Cognitive Services QnA Maker Editor

Ensure that the members of a group named Consultant-Accountants can create and amend the FAQs.

QnA Maker Editor: API permissions:

1. Create KB API

2. Update KB API

3. Replace KB API

4. Replace Alterations

5. " Train API " [in new service model v5]

Box 3: Cognitive Services QnA Maker Read

Ensure that the members of a group named the Agent-CustomerServices can browse the FAQs.

QnA Maker Read: API Permissions:

1. Download KB API

2. List KBs for user API

3. Get Knowledge base details

4. Download Alterations

Generate Answer

Scenario: All planned projects must support English, French, and Portuguese.

If a bot uses multiple LUIS models and QnA Maker knowledge bases (knowledge bases), you can use the Dispatch tool to determine which LUIS model or QnA Maker knowledge base best matches the user input. The dispatch tool does this by creating a single LUIS app to route user input to the correct model.

C: Analyze Receipt - Get Analyze Receipt Result.

Query the status and retrieve the result of an Analyze Receipt operation.

Request URL: https://{endpoint}/formrecognizer/v2.0-preview/prebuilt/receipt/analyzeResults/{resultId}

E: POST {Endpoint}/vision/v3.2/read/analyze

Use this interface to get the result of a Read operation, employing the state-of-the-art Optical Character Recognition (OCR) algorithms optimized for text-heavy documents.

Scenario: Contoso plans to develop a document processing workflow to extract information automatically from PDFs and images of financial documents

The document processing solution must be able to process standardized financial documents that have the following characteristics:

- Contain fewer than 20 pages.

- Be formatted as PDF or JPEG files.

- Have a distinct standard for each office.

*The document processing solution must be able to extract tables and text from the financial documents.

The document processing solution must be able to extract information from receipt images.

The wiki contains text in English, French and Portuguese.

Scenario: All planned projects must support English, French, and Portuguese.

The Document Extraction skill extracts content from a file within the enrichment pipeline. This allows you to take advantage of the document extraction step that normally happens before the skillset execution with files that may be generated by other skills.

Note: The Translator Text API will be used to determine the from language. The Language detection skill is not required.

Scenario: Support autocompletion and autosuggestion based on all product name variants.

A: Call a suggester-enabled query, in the form of a Suggestion request or Autocomplete request, using an API. API usage is illustrated in the following call to the Autocomplete REST API.

POST /indexes/myxboxgames/docs/autocomplete?search & api-version=2020-06-30

{

" search " : " minecraf " ,

" suggesterName " : " sg "

}

B: In Azure Cognitive Search, typeahead or " search-as-you-type " is enabled through a suggester. A suggester provides a list of fields that undergo additional tokenization, generating prefix sequences to support matches on partial terms. For example, a suggester that includes a City field with a value for " Seattle " will have prefix combinations of " sea " , " seat " , " seatt " , and " seattl " to support typeahead.

F. Use the default standard Lucene analyzer ( " analyzer " : null) or a language analyzer (for example, " analyzer " : " en.Microsoft " ) on the field.

Within a search service, synonym maps are a global resource that associate equivalent terms, expanding the scope of a query without the user having to actually provide the term. For example, assuming " dog " , " canine " , and " puppy " are mapped synonyms, a query on " canine " will match on a document containing " dog " .

Create synonyms: A synonym map is an asset that can be created once and used by many indexes.

Comprehensive Detailed Explanation

The requirement is to build an Adaptive Card for the chatbot that supports multilingual product display and meets accessibility requirements.

Product Name Display

The product JSON stores localized names like " name " : { " en " : " ... " , " es " : " ... " , " pt " : " ... " } .

To dynamically render based on the current language, you must use name[language].

This ensures the product name shown in the card adapts automatically to the user’s preferred language.

Stock Level Warning

Business rules require warnings when stock is low or out of stock.

This means you display the stock warning only when stockLevel is not ' OK ' .

Therefore, the correct condition is:

" $when " : " ${stockLevel != ' OK ' } "

Image Alt Text for Accessibility

Accessibility requirements specify that all images must have alt text in English, Spanish, and Portuguese.

In the product JSON, " alttext " is also localized:

" altText " : { " en " : " Bicycle " , " es " : " Bicicleta " , " pt " : " Bicicleta " }

To render the correct localized alt text dynamically, use image.altText[language].

Correct Selections:

First blank: name[language]

Second blank: " $when " : " ${stockLevel != ' OK ' } "

Third blank: image.altText[language]

Microsoft References

Adaptive Cards Templating

Adaptive Card $when property

Multilingual data in Azure Cognitive Search and Adaptive Cards

Box 1: api-nam.cognitive.microsofttranslator.com

https://docs.microsoft.com/en-us/azure/cognitive-services/translator/reference/v3-0-reference

Box 2: /translate

Box 1: QnA Maker Editor

Scenario: Provide all employees with the ability to edit Q & As.

The QnA Maker Editor (read/write) has the following permissions:

Create KB API

Update KB API

Replace KB API

Replace Alterations

" Train API " [in

new service model v5]

Box 2: Contributor

Scenario: Only senior managers must be able to publish updates.

Contributor permission: All except ability to add new members to roles

Scenario: All videos must have transcripts that are associated to the video and included in product descriptions.

Product descriptions, transcripts, and all text must be available in English, Spanish, and Portuguese.

Step 1: Upload the video to blob storage

Given a video or audio file, the file is first dropped into a Blob Storage. T

Step 2: Index the video by using the Video Indexer API.

When a video is indexed, Video Indexer produces the JSON content that contains details of the specified video insights. The insights include: transcripts, OCRs, faces, topics, blocks, etc.

Step 3: Extract the transcript from the Video Indexer API.

Step 4: Translate the transcript by using the Translator API.

Box 1: Azure Blob storage

At the start of the pipeline, you have unstructured text or non-text content (such as images, scanned documents, or JPEG files). Data must exist in an Azure data storage service that can be accessed by an indexer.

Box 2: Computer Vision API

Scenario: Provide users with the ability to search insight gained from the images, manuals, and videos associated with the products.

The Computer Vision Read API is Azure ' s latest OCR technology (learn what ' s new) that extracts printed text (in several languages), handwritten text (English only), digits, and currency symbols from images and multi-page PDF documents.

Box 3: Translator API

Scenario: Product descriptions, transcripts, and all text must be available in English, Spanish, and Portuguese.

Box 4: Azure Files

Scenario: Store all raw insight data that was generated, so the data can be processed later.

Parameter type: stream

Visual features list: VisualFeatureTypes.Description

Result to use: results.Description.Captions[0]

To generate accessible alt text , you should use the caption produced by Azure Computer Vision’s Description feature (it produces a human-readable sentence with a confidence score). Therefore:

The image input should be a stream , because you’re uploading images (not just passing URLs) during product creation and AnalyzeImage…Async supports image streams .

Request the Description feature in VisualFeatureTypes so the service returns results.Description.Captions.

Use results.Description.Captions[0] and return the caption text if its confidence is high enough (e.g., > 0.5) to meet the accessibility requirement that all images must have relevant alt text .

Other features (Tags, Objects, Brands) are useful for enrichment but do not directly return natural-language captions suitable for alt text.

Microsoft Azure AI Solution References

Computer Vision (Image Analysis) – Description/Captions and features: Microsoft Docs, Image Analysis – VisualFeatureTypes.Description returns description.captions.

https://learn.microsoft.com/azure/ai-services/computer-vision/concept-image-analysis

SDK usage (analyze image from a stream): Microsoft Docs, Analyze an image by using the Computer Vision client library .

https://learn.microsoft.com/azure/ai-services/computer-vision/how-to/call-analyze-image?tabs=version-3-2#analyze-an-image-from-a-stream