34 free Microsoft AI-901 practice questions from the CertStash pack of 158, each with the correct answer and a full explanation. The rule is the same for every exam: ten free on the product page, 34 free here, all 158 in the pack, and each step includes the one before it.
Get all 158 questions (US$39) · Download the first 10 as a PDF · Read the study guide
Question 1
You are using the Azure Speech SDK to develop a Python application that supports real-time spoken conversations.
Which Azure Speech class should you use to configure the connection to the Azure Speech service?
Show answer and explanation
Correct answer: D. SpeechConfig
SpeechConfig holds the settings the Speech SDK uses to connect to the Azure Speech service, such as the key and region (or endpoint), along with options like the recognition language or synthesis voice. Every recognizer or synthesizer is created from a SpeechConfig object, so it is the class that configures the service connection for a Python application that handles real-time spoken conversations. The audio classes only describe where sound comes from or goes to, and the synthesizer performs the speech work after the connection is configured.
Why the other options are wrong
- A. AudioOutputConfig only sets where synthesized audio is sent, such as a speaker, file, or stream.
- B. SpeechSynthesizer performs text to speech and receives a SpeechConfig rather than configuring the connection itself.
- C. AudioConfig only selects the audio input or output device or file, not the connection to the service.
Question 2
You need to create an AI agent in Microsoft Foundry that follows a specific role and behavior when responding to users.
What should you configure?
Show answer and explanation
Correct answer: A. system instructions
System instructions tell a Microsoft Foundry agent who it is and how it should behave. They set the agent's role, tone, rules, and boundaries, and the model applies them to every conversation turn, so they are what you configure to make the agent follow a specific role and behavior. The other settings tune randomness, throughput, or response length, but none of them describes how the agent should act or what persona it should adopt.
Why the other options are wrong
- B. Temperature only controls how random or creative the generated text is.
- C. Tokens per minute is a deployment rate limit that governs throughput, not behavior.
- D. Max completion tokens only caps the length of each response.
Question 3
You are developing a web app that processes invoices to calculate expenses.
You need to extract structured fields, including nested values, from the invoices by using a defined schema.
What should you use?
Show answer and explanation
Correct answer: C. an analyzer in Azure Content Understanding in Foundry Tools
An analyzer in Azure Content Understanding in Foundry Tools processes documents and returns structured output that matches a field schema you define. The schema can include simple fields, lists, and nested objects such as line items with their own descriptions, quantities, and amounts, which is exactly what an invoice expense app needs. Content Understanding also includes a prebuilt invoice analyzer that can be used as is or extended with custom fields, so the app receives ready-to-use structured data instead of raw text.
Why the other options are wrong
- A. An OCR-only pipeline returns raw text and layout without mapping it to a schema of named or nested fields.
- B. A Speech transcription workflow converts audio to text and does not process invoice documents.
- D. Azure AI Search indexes and queries content rather than extracting schema-based fields from documents.
Question 4
What are two purposes of instructions when prompting a generative AI model? Each correct answer presents part of the solution.
NOTE: Each correct selection is worth one point.
Show answer and explanation
Correct answer: A, E. defines the agent’s role and behavior; defines constraints on the model’s responses
Instructions, often given as a system message, tell a generative AI model how to act. They define the agent's role and behavior, such as acting as a friendly support assistant with a set tone and style, and they define constraints on the model's responses, such as limiting answers to certain topics, requiring a specific format, or refusing certain requests. Together these shape every response the model produces. Choosing the model, the region where it runs, and its throughput allocation are all deployment decisions made outside the prompt.
Why the other options are wrong
- B. The model is chosen when you create the deployment, not by the instructions in a prompt.
- C. The Azure region is determined by where the resource and deployment are created.
- D. Tokens per minute is a quota assigned to the deployment and is not set by instructions.
Question 5
You need to build an AI solution that produces new product images based on written descriptions provided by users.
Which AI workload should you use?
Show answer and explanation
Correct answer: B. image generation
Image generation is the generative AI workload that creates new images from natural language descriptions. A user writes a prompt describing the product, and a model such as DALL-E or gpt-image-1 produces an original image that matches it. Because the solution must produce new product images from written descriptions rather than interpret existing pictures, image generation is the workload to use. The other options all analyze images that already exist.
Why the other options are wrong
- A. Image analysis describes or tags the content of existing images and does not create new ones.
- C. Object detection locates and labels objects within existing images.
- D. OCR extracts printed or handwritten text from images rather than creating images.
Question 6
You have a Microsoft Foundry project that has a generative AI model deployment.
You need to ensure that responses generated by the model minimize costs and remain within a defined length.
Which parameter should you configure?
Show answer and explanation
Correct answer: C. Max Completion Tokens
Max Completion Tokens sets an upper limit on the number of tokens the model can generate in a response. Because generative AI models are billed by token usage, capping the output length both keeps responses within a defined size and limits the cost of each request. The other settings influence how the model chooses words or which model build is used, but none of them restrict how long a response can be.
Why the other options are wrong
- A. Top P controls the range of likely tokens the model samples from, not the length of the response.
- B. Temperature adjusts the randomness and creativity of the output, not how many tokens are generated.
- D. Model version settings determine which release of the model the deployment uses and do not cap response length or cost per response.
Question 7
Your company processes customer support emails.
You need to implement an AI solution that automatically identifies mentions of people, organizations, and locations in the emails.
Which text analysis technique should you use?
Show answer and explanation
Correct answer: A. Named Entity Recognition (NER)
Named Entity Recognition (NER) identifies and categorizes entities that appear in text, such as people, organizations, locations, dates, and quantities. Running NER over customer support emails automatically finds every mention of a person, company, or place and labels each with its entity category, which is exactly the requirement described.
Why the other options are wrong
- B. Key phrase extraction returns the main talking points of a text but does not classify them as people, organizations, or locations.
- C. Sentiment analysis determines whether text expresses positive, negative, or neutral opinion rather than identifying entities.
- D. Summarization condenses a document into a shorter version and does not label mentions of people, organizations, or locations.
Question 8
You need to build an AI solution that generates marketing email drafts based on a short description of a product and its target audience.
Which AI workload should you use?
Show answer and explanation
Correct answer: C. generative AI
Generative AI creates new original content, such as text, images, or code, in response to a prompt. Producing a marketing email draft from a short description of a product and its target audience is a text generation task, because the model must compose new text rather than analyze or label existing input.
Why the other options are wrong
- A. Speech recognition converts spoken audio into text and does not compose new written content.
- B. Text classification assigns existing text to predefined categories rather than writing new text.
- D. Computer vision interprets images and video, which is not needed to write marketing emails.
Question 9
You are developing an AI-powered customer support application.
Which task is an example of the Microsoft responsible AI principle of inclusiveness?
Show answer and explanation
Correct answer: C. Design the interface to support multiple languages and screen readers.
Inclusiveness means AI systems should empower and engage everyone, including people with disabilities and people from different language and cultural backgrounds. Designing the support application's interface to work in multiple languages and with screen readers removes barriers so that more people can use it, which is a direct application of the inclusiveness principle.
Why the other options are wrong
- A. Encrypting stored data and restricting access with role-based controls is an example of privacy and security.
- B. Evaluating outputs across demographic groups to reduce bias is an example of fairness.
- D. Explaining how predictions are generated is an example of transparency.
Question 10
You have a Microsoft Foundry project that contains a generative AI model deployment.
You test the model by using the Foundry playground.
You need to develop an application that sends requests to the deployed model.
Which information must the application include to call the model?
Show answer and explanation
Correct answer: D. The model endpoint and authentication credentials
An application calls a deployed model by sending requests to the deployment's endpoint and authenticating with either an API key or a Microsoft Entra ID token. The endpoint tells the application where to send the request, and the credentials prove it is authorized to use the deployment. The playground is only a testing tool, and moving to code requires these same connection details.
Why the other options are wrong
- A. Playground session history is not needed to call the deployed model; each request carries its own messages.
- B. The training dataset is not sent when calling a deployed model for inference.
- C. The project display name alone does not identify the deployment endpoint or authorize the request.
What people say about the full AI-901 pack 4.5 from 2
Deeply relieved after passing Microsoft AI-901 exam. Got 95% marks and stunned all my colleagues and company owner too. This is only the results of the 100% real exam questions and passed
Question 11
You have a Python application that extracts fields from invoices by using Azure Content Understanding in Foundry Tools.
You submit a PDF for analysis.
What should the application do to retrieve the results?
Show answer and explanation
Correct answer: C. Poll the operation results until the asynchronous analysis job succeeds.
Azure Content Understanding in Foundry Tools processes an analyze request as a long-running asynchronous operation. Submitting the PDF returns an operation reference (in the Python SDK, begin_analyze() returns a poller), and the application polls that operation until its status reports success. The completed result then contains the extracted content and the invoice fields defined by the analyzer, ready for the application to read.
Why the other options are wrong
- A. Calling begin_analyze() again only starts additional analysis jobs and never makes the operation synchronous.
- B. The analyzer schema describes which fields to extract; downloading it does not return the results of the submitted PDF.
- D. Running separate OCR and mapping fields in code bypasses the service instead of retrieving the analysis it already performed.
Question 12
You have a Microsoft Foundry project that contains a vision-enabled model deployment.
You use the Azure OpenAI Responses API to send a prompt to the model.
You need to provide an image for analysis.
Which content item should you include in the request?
Show answer and explanation
Correct answer: C. input_image
The Azure OpenAI Responses API represents an image supplied by the caller as a content item with the type input_image. The item carries an image_url value that holds either a URL or a base64 data URL, and it is placed alongside input_text items in the user message's content array. This tells the vision-enabled model that the image is input to analyze.
Why the other options are wrong
- A. The image_generation item relates to a tool that creates new images, not to supplying an image for analysis.
- B. output_image is not an input content type, so it cannot carry an image supplied for analysis.
- D. There is no image_base64 content type; base64 data is sent inside an input_image item.
Question 13
You are developing an application that continuously transcribes speech from a default microphone by using Azure Speech in Microsoft Foundry Tools.
You need to perform speech-to-text transcription from the microphone.
Which Python class should you use?
Show answer and explanation
Correct answer: A. SpeechRecognizer
SpeechRecognizer is the Azure Speech SDK class that performs speech to text. You create it from a SpeechConfig and an AudioConfig that uses the default microphone, then call start_continuous_recognition to receive recognized text through events for as long as the user keeps speaking. It is the class that carries out continuous transcription from the microphone.
Why the other options are wrong
- B. SpeechSynthesizer converts text into spoken audio, which is the opposite of transcription.
- C. SpeechConfig only holds settings such as the key, region and language and does not recognize speech on its own.
- D. AudioOutputConfig defines where synthesized audio is played or saved and is used for speech output, not input.
Question 14
You have an Azure subscription.
You need to use Azure Content Understanding in Foundry Tools to extract structured data from invoices.
What should you provision?
Show answer and explanation
Correct answer: C. a Microsoft Foundry resource
Azure Content Understanding runs on a Microsoft Foundry resource, which is the Azure resource that provides the endpoint, keys, and billing for Foundry Tools. After provisioning the Foundry resource, you can call Content Understanding analyzers against its endpoint to extract structured data from invoices. The resource must exist in the subscription before any analyzer can be used.
Why the other options are wrong
- A. A Foundry project is a workspace created inside a Foundry resource, so the resource is what must be provisioned.
- B. An Azure OpenAI resource hosts OpenAI model deployments and does not provide Content Understanding.
- D. Azure AI Search indexes and searches content but does not extract structured fields from invoices.
Question 15
You have a Microsoft Foundry project that contains an agent named Agent1.
You need to ensure that Agent1 always calls an Azure function when the agent responds to user input.
To what should you set tool_choice for Agent1?
Show answer and explanation
Correct answer: A. required
The tool_choice setting controls whether an agent may, must, or must not call its tools. Setting tool_choice to required forces the model to call at least one of its configured tools on every run, so when the Azure function is the agent's tool, Agent1 always invokes it before responding to user input. This guarantees the function call happens instead of leaving the decision to the model.
Why the other options are wrong
- B. The auto setting lets the model decide whether to call a tool, so the Azure function might not be called.
- C. The none setting prevents the agent from calling any tools, so the Azure function would never be called.
Question 16
You are developing an application that processes voicemail recordings by using Azure Content Understanding in Foundry Tools.
Which feature does Azure Content Understanding use to convert audio to text?
Show answer and explanation
Correct answer: B. transcription
Azure Content Understanding processes audio files such as voicemail recordings by first producing a transcription, which converts the spoken audio into text. The transcript, with speaker and timing information, then serves as the basis for extracting fields and generating summaries. Transcription is therefore the feature that turns audio into text in Content Understanding.
Why the other options are wrong
- A. Key phrase extraction identifies important terms in text that already exists and does not convert audio to text.
- C. Voice Live is a real-time speech-to-speech conversational capability, not the Content Understanding feature that converts recordings to text.
- D. Optical character recognition reads printed or handwritten text from images and documents, not spoken audio.
Question 17
Your company has thousands of recorded customer support calls in multiple languages stored as audio files in Azure Storage.
You need to generate text transcripts of all the recordings.
Which Azure Speech in Foundry Tools capability should you use?
Show answer and explanation
Correct answer: A. speech to text batch transcription
Batch transcription in Azure Speech is designed for transcribing large volumes of prerecorded audio stored in Azure Storage. You submit a job that points to the audio files or a storage container, the service processes the recordings asynchronously, and it writes the transcripts back for retrieval. It supports many languages and scales to thousands of files without the application having to stream each one. That makes it the right fit for an archive of recorded customer support calls.
Why the other options are wrong
- B. Real-time transcription streams live or short audio for immediate results and is not intended for processing a large stored archive.
- C. Text to speech converts text into spoken audio, which is the opposite of producing transcripts.
- D. Speech translation converts speech into another language, while the requirement is only to produce transcripts of the recordings.
Question 18
You are developing a lightweight application that will call an agent programmatically by using the Microsoft Foundry SDK.
You need to identify the agent ID.
What should you do?
Show answer and explanation
Correct answer: D. In the Foundry playground, select Code, and view the .env variables.
The agent playground in Microsoft Foundry includes a Code option that shows sample code for calling the agent from an application. Alongside the sample, it lists the environment variables the code expects in a .env file, including the project endpoint and the agent ID. Copying those values gives a lightweight application exactly what it needs to reference the agent through the Microsoft Foundry SDK.
Why the other options are wrong
- A. A model family name identifies the underlying model, not the agent that uses it.
- B. The project name identifies the project that contains the agent, not the agent itself.
- C. A tool name identifies a capability the agent uses and cannot be used to address the agent.
Question 19
You are developing an application that summarizes customer comments by using Azure Language in Foundry Tools.
You need to install the Azure Language SDK.
Which Python package should you install?
Show answer and explanation
Correct answer: B. azure-ai-textanalytics
The Python client library for Azure Language in Foundry Tools text features is distributed as the azure-ai-textanalytics package. Its TextAnalyticsClient provides sentiment analysis, key phrase extraction, entity recognition, and extractive and abstractive summarization, so installing this package gives the application what it needs to summarize customer feedback.
Why the other options are wrong
- A. The azure-ai-translation-text package is the client for Azure Translator and does not provide summarization.
- C. The azure-search-documents package is the client for Azure AI Search indexes and queries.
- D. LangChain is a third-party orchestration framework, not the Azure Language SDK.
Question 20
In the Microsoft Foundry portal, you create an agent named Agent1, and then deploy the agent.
You open the Foundry playground to test Agent1.
What does the playground use to ensure that the tests reflect production behavior?
Show answer and explanation
Correct answer: C. the configuration of Agent1
When you open a deployed agent in the Foundry playground, the playground loads that agent's own configuration: its model deployment, instructions, tools, and knowledge. Every test conversation therefore runs with the same settings the agent uses when applications call it, so the behavior you observe reflects production behavior for Agent1.
Why the other options are wrong
- A. A generic default configuration would ignore Agent1's instructions and tools and would not reflect production.
- B. Testing the base model without instructions or tools shows raw model output, not how Agent1 behaves.
- D. An undeployed draft differs from what is running in production, so testing it cannot confirm production behavior.
Question 21
You are developing an application that analyzes voicemail recordings by using Azure Content Understanding in Foundry Tools.
You need to extract a transcript and structured information from the recordings.
Which type of analyzer should you use?
Show answer and explanation
Correct answer: C. audio analyzer
Azure Content Understanding provides analyzers for each content modality. An audio analyzer accepts audio files such as voicemail recordings, generates a transcript of the speech, and then extracts the structured fields defined in its schema, such as caller name, reason for calling, or callback number. Because the input is a recording and both a transcript and structured data are required, the audio analyzer is the correct choice.
Why the other options are wrong
- A. A document analyzer processes documents such as PDFs and scanned forms, not audio recordings.
- B. A video analyzer processes video files with visual frames and is not intended for audio-only voicemail.
- D. An image analyzer processes still images and cannot transcribe speech.
Question 22
You are developing an application that records a user’s voice and sends the recorded audio to a deployed multimodal model in Microsoft Foundry.
You need to send the user’s request to the model.
What should you include in the request?
Show answer and explanation
Correct answer: C. an audio input prompt
A multimodal model that accepts audio can process speech directly, so the application should send the recorded audio as an audio input in the prompt. The request includes the audio content, typically base64-encoded with its format, as part of the user message, and the model interprets the spoken request itself. This keeps the tone and wording of the user's voice and avoids a separate transcription step.
Why the other options are wrong
- A. Converting audio frames into an image is not how audio is supplied to a model and loses the spoken content.
- B. Sending a transcript would require a separate speech to text step and does not use the model's ability to accept audio.
- D. An embedding vector represents text for similarity search and is not a prompt a model can answer.
Question 23
You have a Microsoft Foundry project that contains a vision-enabled GPT model deployment.
You need to develop an application that sends an image and a text question to the model in a single request.
What should you use?
Show answer and explanation
Correct answer: D. the Azure OpenAI Responses API
The Azure OpenAI Responses API accepts a single request whose input combines several content parts, such as an input_text part holding the question and an input_image part holding the picture, and sends them together to a vision-enabled GPT deployment. The model reasons over both parts at once and returns a text answer that draws on the visual content. This is exactly the pattern needed when an application must ask a question about a picture in one call to a model deployed in a Microsoft Foundry project.
Why the other options are wrong
- A. The Images API generates or edits pictures from prompts; it does not answer a text question about a supplied picture.
- B. The Speech to text REST API transcribes audio and cannot accept a picture or a text question.
- C. The Image Analysis REST API returns predefined visual features such as tags and captions and does not send prompts to a GPT deployment.
Question 24
You have a Microsoft Foundry project that contains a vision-enabled model deployment.
You are developing an application that sends images to the model.
You need to ensure that the model can analyze the images.
In which two formats can you provide the images? Each correct answer presents part of the solution.
NOTE: Each correct selection is worth one point.
Show answer and explanation
Correct answer: A, D. a base64-encoded image data string; a publicly accessible URL of the image
Vision-enabled models in Microsoft Foundry accept image input in two forms. The image can be embedded directly in the request as base64-encoded image data, typically formatted as a data URL that states the image type, which suits local files. Alternatively, the request can contain a publicly accessible URL that the service retrieves the image from, which suits images already hosted online. In both cases the model receives the actual pixels and can analyze the image content.
Why the other options are wrong
- B. A JSON description of the image gives the model text about the image, not the image itself.
- C. A text description of the image contains no pixel data, so the model has no image to analyze.
Question 25
You need to compare the costs of large language models (LLMs) for a generative AI solution.
What should you use in the Microsoft Foundry portal?
Show answer and explanation
Correct answer: D. Model leaderboard
Model leaderboards in the Microsoft Foundry portal rank and compare models on quality, safety, estimated cost and throughput. The cost view shows the relative price of each model so you can weigh it against quality scores and choose a model that fits your budget. You can also select models to compare side by side before deploying one, which directly supports comparing the costs of large language models for a generative AI solution.
Why the other options are wrong
- A. Compliance relates to governance and policy settings, not to comparing model prices.
- B. The evaluator catalog lists evaluators for scoring model outputs and does not compare model costs.
- C. Tools are capabilities you attach to agents, such as search or code execution, and do not compare models.
Question 26
You are developing an application that uses the Azure Language SDK and a Microsoft Foundry resource.
You need to return a list of important topics from unstructured text.
Which Python method should you use?
Show answer and explanation
Correct answer: B. client.extract_key_phrases()
Key phrase extraction in Azure Language identifies the main talking points in unstructured text and returns them as a list of phrases. In the Python SDK, the TextAnalyticsClient method extract_key_phrases takes a batch of documents and returns the key phrases found in each one. A list of the important topics in the text is exactly what this method produces.
Why the other options are wrong
- A. begin_extract_summary produces summary sentences drawn from the text, not a list of topics.
- C. recognize_linked_entities identifies well-known entities and links them to a knowledge base such as Wikipedia rather than listing topics.
- D. recognize_entities returns named entities such as people, places and organizations, not the main topics of the text.
Question 27
You are developing an application that analyzes uploaded video files by using Azure Content Understanding in Foundry Tools.
You need to ensure that the application extracts structured information from the videos.
What should the application call?
Show answer and explanation
Correct answer: B. the Analyze API in Content Understanding
Azure Content Understanding extracts structured information from documents, images, audio and video by running an analyzer over the content. The application calls the Analyze API with the video file and an analyzer, either prebuilt or custom, and receives structured output such as transcripts, scene and shot descriptions, and the fields defined in the analyzer schema. This is the call that turns an uploaded video into structured data the application can use.
Why the other options are wrong
- A. Sentiment analysis in Azure Language evaluates the opinion expressed in text and does not process video files.
- C. Image Analysis in Azure Vision works on still images and does not extract structured information across a video.
- D. Azure OpenAI provides generative models and is not the Content Understanding operation that analyzes video into structured fields.
Question 28
You have a Microsoft Foundry project that contains a vision-enabled model deployment.
You need to develop an application that sends a message containing text and an image URL. The solution must ensure the quickest response time.
Which message structure should you include in the request?
Show answer and explanation
Correct answer: D. a user message that includes both a text item and an image item in the content array
A vision-enabled chat model accepts a user message whose content array holds several items, for example one text item with the question and one image item with the image URL. Sending both items together in a single user message lets the model process the text and the image in one call, which avoids extra round trips and gives the quickest response. The user role is the correct place for the content the user wants analyzed.
Why the other options are wrong
- A. A system message sets the model's behavior and instructions, so it is not where the user's image and question belong.
- B. Using a system message and splitting the image into a second request adds a round trip and slows the response.
- C. Sending the image in a separate request adds a round trip and separates the image from the question it relates to.
Question 29
You are developing an application that extracts fields from PDFs by using Azure Content Understanding in Foundry Tools.
You need to use the Python SDK to submit a PDF for analysis and retrieve the extraction results.
What should you do?
Show answer and explanation
Correct answer: C. Call begin_analyze(), and then call poller.result() to retrieve the results.
Analysis in Azure Content Understanding is a long-running operation. In the Python SDK, begin_analyze() submits the PDF to the chosen analyzer and returns a poller object that tracks the operation. Calling poller.result() waits until the analysis completes and returns the result containing the extracted fields and their values. This asynchronous pattern is how the SDK submits content and retrieves the extraction results.
Why the other options are wrong
- A. Analysis runs as a long-running operation, so the SDK does not return extracted fields synchronously in the same request.
- B. Extraction results are returned in the body of the completed operation result, not in request headers.
- D. Running OCR and mapping fields in custom code bypasses the analyzer and its schema, which is the purpose of Content Understanding.
Question 30
You are developing an application named App1 that analyzes customer feedback.
App1 will call Azure Language in Foundry Tools from a generative AI agent in Microsoft Foundry.
You need to ensure that the agent can access Azure Language text analysis.
What should you configure?
Show answer and explanation
Correct answer: B. the Azure Language MCP server
A generative AI agent in Microsoft Foundry calls outside services through tools, and the Azure Language MCP server exposes Azure Language in Foundry Tools operations, such as sentiment analysis, key phrase extraction, and entity recognition, as Model Context Protocol tools. Adding this server to the agent lets App1's agent discover and invoke the text analysis operations it needs to analyze customer feedback, without writing custom integration code for each operation.
Why the other options are wrong
- A. A model deployment supplies the agent's reasoning model but does not give it access to Azure Language text analysis.
- C. An Azure Speech endpoint provides speech recognition and synthesis, not text analysis of written feedback.
- D. An Azure AI Search index supports document retrieval and grounding rather than calling Azure Language text analysis.
Question 31
You have a Microsoft Foundry project that contains a vision-enabled model deployment.
You need to create a prompt that ensures the model produces a relevant and useful response.
What should you include in the prompt?
Show answer and explanation
Correct answer: B. a description of the task
A vision-enabled model responds best when the prompt clearly describes the task it should perform with the image, such as describing a scene, extracting text, or answering a specific question about what is shown. A clear task description gives the model context and direction, which leads to relevant and useful output instead of a generic description. Prompt engineering guidance for vision models stresses stating the task, the desired detail and the output format.
Why the other options are wrong
- A. The model version is a deployment detail and does not help the model understand what to do with the image.
- C. The deployment name is used by the code to route the request and adds nothing useful to the prompt content.
- D. A dataset ID is not something the model can interpret, and images are supplied directly as content in the request.
Question 32
You are developing a simple application that uses the Microsoft Foundry SDK to send chat prompts to a model.
Which three elements should you provide in the code to run the application? Each correct answer presents part of the solution.
NOTE: Each correct selection is worth one point.
Show answer and explanation
Correct answer: A, D, E. the model deployment name; Microsoft Entra ID credentials; the endpoint URL
A Microsoft Foundry SDK application needs three things to send chat prompts. The endpoint URL tells the client which Foundry project or resource to connect to. Microsoft Entra ID credentials, typically supplied through DefaultAzureCredential, authenticate the application so the request is authorized. The model deployment name identifies which deployed model receives the prompt, and it is passed in the model parameter of each chat request. With these three values the client can connect, authenticate and route the conversation to the correct model.
Why the other options are wrong
- B. The chat method itself defines the operation, so no separate task type needs to be supplied.
- C. The display name is a portal label and is not used by the SDK to connect, which relies on the endpoint URL instead.
Question 33
You have a Microsoft Foundry project that contains a vision-enabled chat model deployment.
You are developing a Python application that sends a request containing a user prompt and a local JPEG image.
You need to include the image as binary data in the request.
To what should you set the image_url.url field?
Show answer and explanation
Correct answer: A. data:image/jpeg;base64,<Base64-encoded image data>
To include a local JPEG as binary data, the application reads the file, encodes the bytes as base64, and sets the image_url.url field to a data URL that begins with data:image/jpeg;base64, followed by the encoded string. The data URL embeds the image itself in the request and declares its media type, so the vision-enabled model receives the picture without needing to fetch it from anywhere.
Why the other options are wrong
- B. A file URI points to a path on the local machine, which the model service cannot access.
- C. A Windows file path is only meaningful on the local computer and does not send the image data.
- D. A Blob Storage URL references a hosted file that the service must retrieve, rather than embedding the image as binary data.
Question 34
You need to convert written customer notifications into natural-sounding spoken audio that can be played over a phone system.
Which Azure Speech in Foundry Tools capability should you use?
Show answer and explanation
Correct answer: B. speech synthesis
Speech synthesis, also called text to speech, converts written text into spoken audio. Azure Speech in Foundry Tools offers neural voices that produce natural-sounding speech, and the output can be generated in formats suitable for telephony. This makes speech synthesis the capability for turning written customer notifications into audio played over a phone system.
Why the other options are wrong
- A. Speech recognition converts spoken audio into text, which is the opposite direction.
- C. Speaker recognition identifies or verifies who is speaking and does not generate audio.
- D. Speech translation converts speech from one language to another rather than turning written text into speech.
That was 34 of 158.
The full Microsoft AI-901 pack has all 158 questions, each with the answer, the explanation and why the other options are wrong, plus a questions-only copy for timed runs. US$39, paid once, with free monthly updates and a pass-or-your-money-back guarantee.
Not ready yet? Get the first 28 free as one PDF by email. Get the Half Pack →