Q001 - Question
An architect is documenting which SDK a team should use for each part of a Microsoft Foundry chat solution. Which two statements correctly describe the SDK options? Each correct answer presents a complete solution.
Domain: Implement generative AI and agentic solutions (30–35%) Type: Multiple choice
- A. The OpenAI SDK provides access to project-level features such as agents, evaluations, tracing, and connections.
- B. The Foundry SDK provides access to project-level features such as agents, evaluations, tracing, and connections.
- C. Using the Foundry SDK removes the need to deploy a generative model in the Microsoft Foundry project.
- D. The OpenAI SDK provides model inference with full OpenAI API compatibility.
B and D are correct.
Explanation: The Foundry SDK gives you project-level features such as agents, evaluations, tracing, and connections, while the OpenAI SDK provides model inference with full OpenAI API compatibility. Choose the SDK based on which of these capabilities the application needs.
A is incorrect: Project-level features such as agents, evaluations, tracing, and connections come from the Foundry SDK, not the OpenAI SDK.
C is incorrect: You still deploy a generative model in the Microsoft Foundry project before an application can chat with it.
Q002 - Question
An analytics team wants a generative AI app to calculate summary statistics from an uploaded data file and produce a chart. They plan to enable the code_interpreter tool. Which statement should shape their design?
Domain: Implement generative AI and agentic solutions (30–35%) Type: Single choice
- A. The tool runs the generated code directly on the client computer that submitted the prompt.
- B. The tool can call any internal REST API during code execution to enrich the data set.
- C. The tool only returns Python code as text, so the application must run the code itself.
- D. The tool runs the generated Python code in a sandboxed runtime that has network, timeout, and memory constraints.
D is correct.
Explanation: The code_interpreter tool gives the model a sandboxed Python runtime for tasks such as data analysis and file handling, and that runtime is subject to network, timeout, and memory constraints, so plan work that fits those limits.
A is incorrect: Execution happens in the sandboxed runtime provided by the tool, not on the client computer.
B is incorrect: The sandbox applies network constraints, so you cannot assume the generated code can reach internal APIs.
C is incorrect: The tool generates and runs Python code, so your application does not need to execute the code itself.
Q003 - Question
A retail company wants an application that transcribes recorded customer calls, summarizes them, and extracts product names from scanned receipts. Before any resources are provisioned, what should the solution architect do first?
Domain: Plan and manage an Azure AI solution (25–30%) Type: Single choice
- A. Provision one of every available Azure AI resource so that all options remain open during development.
- B. Write the custom application code first and decide which AI capabilities to add after testing.
- C. Determine the specific AI capabilities the application needs, such as computer speech, natural language processing, and information extraction.
- D. Select the developer tools and SDKs for the project team before defining the application workload.
C is correct.
Explanation: Determining the specific AI capabilities to include, such as generative AI and agents, natural language processing, computer speech, computer vision, and information extraction, helps identify the most appropriate AI services to provision, configure, and use.
A is incorrect: Provisioning every resource does not identify which capabilities the workload requires, and it adds resources you do not need to configure or manage.
B is incorrect: Custom code is only one part of an AI solution, which also combines machine learning models, AI services, and prompt engineering, so the required capabilities should shape the code.
D is incorrect: Tool and SDK selection follows from the capabilities and services you plan to use, not the other way around.
Q004 - Question
A development team needs a single place to organize the models, agents, tools, and knowledge connections for a new AI solution, and the team wants both portal-based and code-based management. What should the team use?
Domain: Plan and manage an Azure AI solution (25–30%) Type: Single choice
- A. A Microsoft Foundry project, managed through the Foundry portal or the SDK.
- B. A separate resource for each model and agent, tracked in a spreadsheet by the team.
- C. A source control repository as the only record of the solution's models and agents.
- D. Prompt engineering documentation stored alongside the application code.
A is correct.
Explanation: Microsoft Foundry projects organize the resources, data, code, and assets of an AI solution, and developers use projects to manage models, agents, tools, and knowledge. You can work with a project through the Foundry portal or through the SDK.
B is incorrect: Manual tracking outside the platform does not provide the project-level access to models, agents, tools, and knowledge connections.
C is incorrect: A repository stores code and files, but it does not provide the project structure that organizes Foundry assets.
D is incorrect: Prompt engineering is one part of an AI solution, and documentation alone does not manage the solution's models, agents, and knowledge connections.
Q005 - Question
A bank plans an AI solution that helps assess loan applications. Governance reviewers ask who answers for the decision logic and for the validation of the trained model. Which responsible AI principle addresses this requirement?
Domain: Plan and manage an Azure AI solution (25–30%) Type: Single choice
- A. Inclusiveness
- B. Accountability
- C. Reliability and safety
- D. Privacy and security
B is correct.
Explanation: Accountability means people are answerable for AI systems, including the developers who train and validate models and define decision logic, and that the work stays within governance principles that meet responsibility and legal standards.
A is incorrect: Inclusiveness concerns designing solutions that serve people across different groups, not assigning responsibility for decision logic.
C is incorrect: Reliability and safety concerns consistent and safe system behavior, not identifying who answers for model validation.
D is incorrect: Privacy and security concerns protecting data used by the solution, not the ownership of decision logic.
Q006 - Question
An insurance company must extract fields from scanned claim forms and translate the extracted text. The team has no data science staff and wants to avoid training models. What approach meets the requirement?
Domain: Plan and manage an Azure AI solution (25–30%) Type: Single choice
- A. Train a custom machine learning model for document processing and a second model for translation.
- B. Build prompt engineering solutions in custom code to replace document processing and translation.
- C. Create a separate Foundry project for each document type to avoid integration work.
- D. Use Foundry Tools, which provide prebuilt APIs and models for tasks such as document intelligence and translation.
D is correct.
Explanation: Foundry Tools are out-of-the-box prebuilt APIs and models that you can integrate into applications for tasks including text analysis, speech, translation, document intelligence, and content understanding. Applications connect to tool-specific endpoints using project authentication or token-based authentication.
A is incorrect: Training custom models requires data science effort that the team does not have, and prebuilt APIs already cover document intelligence and translation.
B is incorrect: Prompt engineering is a separate part of an AI solution and does not replace the prebuilt document intelligence and translation APIs.
C is incorrect: A project organizes solution assets, but creating multiple projects does not by itself provide document extraction or translation capabilities.
Q007 - Question
A team must shortlist models in the Microsoft Foundry portal for a summarization feature in a regulated industry. They want to narrow the catalog to a manageable set before any testing begins. Which approach in the Foundry Models catalog best supports this first step?
Domain: Plan and manage an Azure AI solution (25–30%) Type: Single choice
- A. Filter the catalog by collection, inference task, fine-tuning method, and industry, then review the model cards of the remaining models.
- B. Deploy every model that mentions summarization and compare the responses in the playground.
- C. Sort the catalog alphabetically by provider and select the first model from each provider.
- D. Request a custom model build because the catalog does not expose responsible AI information.
A is correct.
Explanation: The Foundry Models catalog supports searching and filtering by collection, capabilities, source, inference tasks, fine-tuning methods, and industry. After filtering, each model card shows the provider, capabilities, benchmark metrics, responsible AI considerations, and deployment options, so you can shortlist candidates before testing.
B is incorrect: Deploying every candidate model adds deployment work before you have used the filters and model cards to reduce the candidate list.
C is incorrect: Provider name and alphabetical order are not selection criteria; the catalog filters and model card details are the documented way to narrow the list.
D is incorrect: Model cards already include responsible AI considerations, so you do not need a custom build to see that information.
Q008 - Question
An application must call agents and Foundry IQ knowledge stores in a Microsoft Foundry project, and it must also call a prebuilt text analysis API. Which two programmatic interfaces should the developers use? Each correct answer presents part of the solution.
Domain: Plan and manage an Azure AI solution (25–30%) Type: Multiple choice
- A. Visual Studio Code
- B. The Microsoft Foundry SDK
- C. GitHub
- D. Foundry Tools SDKs and REST APIs
B and D are correct.
Explanation: The Microsoft Foundry SDK connects to Foundry projects and Foundry-specific assets, including agents and Foundry IQ knowledge stores. Foundry Tools SDKs and REST APIs give the application access to the prebuilt APIs and models, such as text analysis.
A is incorrect: Visual Studio Code is a development tool that you write and test code in, not an interface that the application uses to call project assets or prebuilt APIs.
C is incorrect: GitHub is a development tool for source code, and it does not provide the application's connection to agents, knowledge stores, or prebuilt APIs.
Q009 - Question
A solution architect must reduce the chance that a generative AI application returns harmful content. The architect wants a mitigation strategy that does not depend on a single control. Which approach matches the layered mitigation model?
Domain: Plan and manage an Azure AI solution (25–30%) Type: Single choice
- A. Rely only on guardrails in the safety system layer, because they inspect every request and response.
- B. Rely only on a system message and trusted grounding data, because they control the content the model uses.
- C. Rely only on the user experience layer, because clear documentation sets user expectations.
- D. Apply mitigations across the model, safety system, system message and grounding, and user experience layers.
D is correct.
Explanation: Mitigation techniques can be applied at four layers: the model, the safety system, the system message and grounding, and the user experience. Combining model selection or fine-tuning, guardrails, trusted grounding data, and input or output controls reduces the likelihood of harmful output more than any single layer does.
A is incorrect: Guardrails address only the safety system layer and leave the model choice, grounding data, and user experience unmitigated.
B is incorrect: A system message and grounding data shape model responses but do not add guardrail checks or user experience controls such as input and output constraints.
C is incorrect: User experience measures such as documentation and input or output controls do not change model behavior or inspect generated content.
Q010 - Question
A news monitoring app must answer questions about events that occurred after the model's training data was collected. The team enables the web_search tool. What else should the team plan for?
Domain: Implement generative AI and agentic solutions (30–35%) Type: Single choice
- A. Review the retrieved sources and results for quality before the content is presented as authoritative.
- B. Upload the same web pages as documents so the model has a second copy of the content.
- C. Retrain the model after each search so the new information becomes part of its training data.
- D. Disable all other tools, because web_search cannot be configured alongside them.
A is correct.
Explanation: The web_search tool retrieves current web information while the response is generated, but the retrieved sources and results still require review for quality before you treat the answer as reliable.
B is incorrect: Duplicating the content as uploaded documents is unnecessary work and does not address the need to check the quality of retrieved results.
C is incorrect: The tool grounds a response in timely external content at request time; it does not add information to the model's training data.
D is incorrect: You can specify multiple tools in a request and let the model decide which one to use.
Q011 - Question
You are starting a new Microsoft Foundry chat application that supports multi-turn conversations. You want the service to track conversation state instead of building that logic in your application code. Which API should you use?
Domain: Implement generative AI and agentic solutions (30–35%) Type: Single choice
- A. The ChatCompletions API, because it stores conversation state in the project endpoint.
- B. The Model playground code sample, called directly as the production API.
- C. The azure-ai-projects package alone, without a chat API.
- D. The Responses API, because it provides stateful multi-turn response generation and is the recommended approach for new applications.
D is correct.
Explanation: The Responses API provides stateful multi-turn response generation through an OpenAI-compatible client, and it is the recommended approach for generating AI responses in new Microsoft Foundry applications.
A is incorrect: The ChatCompletions API does not provide stateful response tracking, so your application must manage the conversation history.
B is incorrect: Playground code samples are a starting point for development and are not a separate API.
C is incorrect: Project-level SDK access does not by itself generate chat responses, so you still choose a chat API.
Q012 - Question
A development team is starting work on a generative AI application in Microsoft Foundry. The compliance lead asks the team to follow a documented, repeatable process for responsible AI that also aligns with the NIST AI Risk Management Framework. Which sequence of stages should the team plan to follow?
Domain: Plan and manage an Azure AI solution (25–30%) Type: Single choice
- A. Deploy the model, collect user feedback, retrain the model, and publish release notes.
- B. Select a model, fine-tune it, add grounding data, and design the user experience.
- C. Map potential harms, measure the harms, mitigate the harms, and manage the solution through deployment and operations.
- D. Run a security review, run a privacy review, run an accessibility review, and run a legal review.
C is correct.
Explanation: The practical process for responsible generative AI has four stages: map potential harms, measure those harms, mitigate them at multiple layers, and manage the solution through deployment and operations. These stages align closely with the NIST AI Risk Management Framework.
A is incorrect: Deployment and feedback activities belong to the manage stage only, so this sequence skips identifying, measuring, and mitigating harms before release.
B is incorrect: Model selection, fine-tuning, grounding data, and user experience design are mitigation techniques applied at different layers, not the overall planning process.
D is incorrect: Legal, privacy, security, and accessibility reviews are prerelease tasks within the manage stage, so they do not replace mapping, measuring, and mitigating harms.
Q013 - Question
A development team deploys a base language model for an internal summarization assistant. Responses are inconsistent in quality and relevance. The team has no labeled training data, no budget for additional infrastructure, and needs to improve results in the current sprint. Which optimization approach fits these constraints?
Domain: Plan and manage an Azure AI solution (25–30%) Type: Single choice
- A. Collect a task-specific dataset and fine-tune the deployed model.
- B. Refine the prompts, including the system message, to improve quality, accuracy, and relevance.
- C. Build a retrieval pipeline over an indexed data source before changing the prompts.
- D. Deploy a second model and route requests between the two deployments.
B is correct.
Explanation: Prompt engineering is the process of designing and refining prompts, including the system message, to improve response quality, accuracy, and relevance. It requires no extra infrastructure and no training data, so it is the approach that matches the team's constraints and timeline.
A is incorrect: Fine-tuning requires a smaller task-specific dataset, which the team does not have.
C is incorrect: Retrieval Augmented Generation adds a retrieval step over a data source, which requires additional setup, and prompt engineering is the starting point for optimization.
D is incorrect: Adding a second deployment does not address prompt quality and adds infrastructure the team cannot fund.
Q014 - Question
You add an agent to a Foundry workflow to classify support tickets. A later flow control node must branch on the classification result. What should you configure so the later node can use the result reliably?
Domain: Implement generative AI and agentic solutions (30–35%) Type: Single choice
- A. Place the agent in a group chat node so other agents can read its reply
- B. Add a second Invoke agent node that repeats the classification for each branch
- C. Store the full free-form text of the agent reply and parse it in the chat node
- D. Configure the Invoke agent node to produce a structured response and store it in a variable
D is correct.
Explanation: Agents are added through an Invoke agent node, and structured responses provide predictable data that can be stored in variables and used by later workflow steps.
A is incorrect: A group chat node coordinates agent interaction and does not by itself supply predictable data for a branch condition.
B is incorrect: Repeating the classification in another Invoke agent node adds cost and can return a different result instead of reusing one stored value.
C is incorrect: Free-form text is not predictable data, so parsing it is less reliable than configuring a structured response.
Q015 - Question
An operations group plans to publish new tools for an agent solution built on Foundry Agent Service every few weeks. Agent downtime is not acceptable, and the group wants to avoid redeploying agents for each tool change. Which outcome should they expect from using MCP for tool integration?
Domain: Implement generative AI and agentic solutions (30–35%) Type: Single choice
- A. Agents call the tools without any client or tool connection because the model selects tools directly.
- B. Agents discover and register the available tools at runtime, so new tools can be used without hardcoding APIs or redeploying the agents.
- C. Agents cache a fixed tool list at deployment time and use that list until the next deployment.
- D. Agents require a separate deployment for each tool so that tools remain isolated from each other.
B is correct.
Explanation: MCP lets Foundry Agent Service discover and register external tools at runtime, so the solution adapts as tools change without hardcoded APIs or agent redeployment.
A is incorrect: The agent still connects to the MCP server through an MCP client or an MCP tool connection before tools can be called.
C is incorrect: A fixed list captured at deployment time is the hardcoded pattern that runtime discovery replaces.
D is incorrect: Tools are hosted on the MCP server and discovered on demand, so each tool does not need its own agent deployment.
Q016 - Question
A team is designing an invoice-processing workflow in Microsoft Foundry. Each step must run in a fixed order, and the output of one agent is always passed to the next agent. No reviewer input is required, and no agents need to discuss results with each other. Which workflow pattern fits this requirement?
Domain: Implement generative AI and agentic solutions (30–35%) Type: Single choice
- A. A sequential workflow
- B. A human-in-the-loop workflow
- C. A group chat workflow
- D. A workflow that runs every agent in parallel and merges results at the end
A is correct.
Explanation: A sequential workflow follows a fixed, step-by-step path, so each agent runs in a defined order and passes its result to the next step.
B is incorrect: A human-in-the-loop workflow pauses for human input or escalation, which this process does not need.
C is incorrect: A group chat workflow coordinates multiple agents interacting on the same task, which adds interaction that this fixed order does not require.
D is incorrect: Parallel execution with a merge step is not the pattern described for a fixed, step-by-step path.
Q017 - Question
A support agent answers policy questions using only the language model's training data. Employees report that the answers are out of date and contain no indication of where the information came from. Which approach addresses both problems?
Domain: Implement generative AI and agentic solutions (30–35%) Type: Single choice
- A. Lower the temperature setting so the model produces more deterministic wording.
- B. Increase the maximum response length so the model can add more detail to each answer.
- C. Use retrieval augmented generation to retrieve relevant organizational content, add it to the query context, and generate a grounded response.
- D. Add a system message that tells the model to answer only when it is confident about the policy.
C is correct.
Explanation: Retrieval augmented generation follows a retrieve, augment, generate pattern. It pulls current organizational content, adds that content to the query context, and produces a response that is grounded in the retrieved sources, which also gives you source transparency.
A is incorrect: A lower temperature changes response variability, but the agent still answers from static training data with no retrieved sources.
B is incorrect: A longer response adds detail from the same outdated training data and does not supply current content or citations.
D is incorrect: Instructing the model to answer only when confident does not give it access to current organizational information or identify the source of an answer.
Q018 - Question
Several developers maintain a shared Foundry workflow. They need a record of each change and a way to manage the workflow definition in source control alongside application code. Which two approaches meet these needs? Each correct answer presents part of the solution.
Domain: Implement generative AI and agentic solutions (30–35%) Type: Multiple choice
- A. Export the workflow canvas as an image and attach it to each pull request
- B. Save the workflow so that Foundry creates a new immutable version for the change
- C. Keep one draft open in the visual designer and have all developers edit it at the same time
- D. Use the YAML representation, which stays synchronized with the visual canvas, for source control and advanced editing
B and D are correct.
Explanation: Every time a workflow is saved, Foundry creates a new immutable version, and the visual canvas and YAML representations stay synchronized so teams can use YAML for advanced editing and source control.
A is incorrect: An image of the canvas is not a workflow definition and cannot be diffed or restored as a version.
C is incorrect: Editing a single open draft does not produce the version history the team needs and offers no review point before changes take effect.
Q019 - Question
A development team plans to build several agents that use different model providers. The team wants the agents to expose consistent capabilities so that developers do not learn a different programming model for each provider. Which characteristic of Microsoft Agent Framework supports this requirement?
Domain: Implement generative AI and agentic solutions (30–35%) Type: Single choice
- A. Every agent is derived from a unified Agent base class.
- B. Each model provider requires a separate agent abstraction.
- C. Telemetry replaces the need for a shared agent abstraction.
- D. Graph-based workflows remove the need to define agents.
A is correct.
Explanation: Microsoft Agent Framework derives every agent from a unified Agent base class, which provides consistent capabilities across different model providers.
B is incorrect: The unified base class prevents the need for a separate agent abstraction for each model provider.
C is incorrect: Telemetry is an enterprise feature for observability, not a replacement for a shared agent abstraction.
D is incorrect: Graph-based workflows coordinate work but do not remove the need to define agents.
Q020 - Question
You complete a hands-on lab in which you create an agent in code and connect a tool definition to a custom tool function. You verified that the agent calls the tool. What should you do next to control cost in the Azure subscription?
Domain: Implement generative AI and agentic solutions (30–35%) Type: Single choice
- A. Detach the tool definition from the agent and keep the deployed resources.
- B. Delete the Azure resources that you created for the lab.
- C. Move the agent and its tool to a different Azure region.
- D. Convert the custom tool function into an OpenAPI specification tool.
B is correct.
Explanation: After you finish the exercise and confirm that the agent calls the custom tool, delete the Azure resources that you created so that they no longer incur cost.
A is incorrect: The deployed resources remain in the subscription, so detaching the tool definition does not address cost.
C is incorrect: Moving resources to another region keeps them deployed and does not remove them.
D is incorrect: Changing the tool implementation type is a design decision and does not clean up the lab resources.
Q021 - Question
You are building an A2A client for a long-running agent task. Users must see partial output as the agent works instead of waiting for the full response. Which client behavior meets this requirement?
Domain: Implement generative AI and agentic solutions (30–35%) Type: Single choice
- A. Send a non-streaming request and poll the Agent Card until the response is complete.
- B. Send a streaming request so that the client receives incremental results.
- C. Send a non-streaming request and let the request handler write partial output to the task store.
- D. Send the request directly to the Agent Executor and bypass the A2A server.
B is correct.
Explanation: The evidence supports B because clients can use streaming requests for incremental results, which meets the requirement to show partial output while the task runs.
A is incorrect: Non-streaming requests return the complete response, and polling the Agent Card only retrieves discovery metadata.
C is incorrect: The task store supports the server's request handling and is not a mechanism for delivering incremental output to the client.
D is incorrect: Clients interact with the hosted A2A server over HTTP; the executor is invoked as part of server-side request processing.
Q022 - Question
A team connects an agent in Foundry Agent Service to a remote MCP server by using an MCP tool object. The team must limit the agent to a subset of the server's tools and must keep a person in the loop before any tool call runs. Which two configuration options of the MCP tool connection meet these requirements? Each correct answer presents part of the solution.
Domain: Implement generative AI and agentic solutions (30–35%) Type: Multiple choice
- A. Open a manual MCP client session for every tool call made by the agent.
- B. Restrict the connection to a list of allowed tools.
- C. Rewrite the tool definitions that are hosted on the remote MCP server.
- D. Require human approval for tool calls.
B and D are correct.
Explanation: When Foundry Agent Service connects to a remote MCP server through an MCP tool object, the connection can restrict which tools are allowed and can require human approval before tool calls run.
A is incorrect: Using the MCP tool object avoids manual MCP client sessions and function-tool wrappers, so adding manual sessions works against that approach.
C is incorrect: Tool definitions remain on the remote MCP server. The connection settings control access and approval rather than changing the hosted definitions.
Q023 - Question
You are developing an AI agent by using the Microsoft Agent Framework. You need to ensure the agent can maintain persistent multi-turn conversations with users. What should you configure to manage the conversation state?
Domain: Implement generative AI and agentic solutions (30–35%) Type: Single choice
- A. A session that uses Foundry service-side history
- B. A custom Python function tool
- C. An execution filter
- D. A graph-based workflow
A is correct.
Explanation: The session acts as the container for the conversation state and can use Foundry service-side history for persistent multi-turn conversations.
B is incorrect: A custom Python function tool allows the agent to take actions, not manage conversation state.
C is incorrect: An execution filter is an enterprise feature for intercepting execution, not for storing conversation history.
D is incorrect: Graph-based workflows coordinate complex agent interactions, but the session itself holds the conversation state.
Q024 - Question
Before releasing an agent that uses several tools, you run a set of common, single-turn requests and verify the answers. You need to complete the test plan to ensure the agent behaves reliably. Which additional testing should you perform?
Domain: Implement generative AI and agentic solutions (30–35%) Type: Single choice
- A. Test edge cases, boundary conditions, multi-turn context, and tool invocation.
- B. Repeat the same common requests with a different model temperature only.
- C. Validate only that the deployed endpoint returns a response for each request.
- D. Replace tool calls with static sample responses during all testing.
A is correct.
Explanation: Thorough testing covers common requests, edge cases, boundaries, multi-turn context, and tool invocation. Testing these areas shows whether the agent keeps context across turns and calls the correct tools before you deploy it.
B is incorrect: Repeating the same requests does not test edge cases, boundaries, multi-turn context, or tool invocation.
C is incorrect: Confirming that the endpoint responds shows availability, but it does not verify the content or the tool behavior of the responses.
D is incorrect: Replacing tool calls with static responses removes tool invocation from the test, which is one of the behaviors you need to verify.
Q025 - Question
A development team must decide between direct portal publishing and Microsoft 365 Agents Toolkit for a Foundry agent. Which two requirements justify using Agents Toolkit? Each correct answer presents a complete solution.
Domain: Implement generative AI and agentic solutions (30–35%) Type: Multiple choice
- A. The agent must be discoverable by every user in the organization.
- B. The agent requires custom single sign-on handling between Teams and the Foundry agent.
- C. The agent must be deployed across multiple environments with advanced debugging.
- D. The agent must be published with a Teams app package that includes name, description, and icons.
B and C are correct.
Explanation: Agents Toolkit creates a proxy application between Teams or Copilot and the Foundry agent. Reserve it for requirements such as custom SSO, middleware, multi-environment deployment, and advanced debugging.
A is incorrect: Organization-wide availability is handled by selecting the organization publishing scope during portal publishing.
D is incorrect: The portal publishing flow already completes Teams metadata and generates the Microsoft 365 package.
Q026 - Question
You publish a Foundry agent to Microsoft Teams from the Foundry portal. In Foundry, the agent successfully called an Azure tool, but in Teams the same tool call fails with an authorization error. What should you do first?
Domain: Implement generative AI and agentic solutions (30–35%) Type: Single choice
- A. Assign the appropriate RBAC permissions on the tool resources to the published agent identity.
- B. Change the publishing scope from organization to shared and publish the agent again.
- C. Rebuild the agent with Microsoft 365 Agents Toolkit so that it can call Azure tools.
- D. Remove the Azure Bot Service resource and call the agent endpoint directly from Teams.
A is correct.
Explanation: The published agent runs under its own identity, so permissions that worked during development do not carry over. Grant that published identity the RBAC permissions it needs on each tool resource after publishing.
B is incorrect: The publishing scope controls who can install and use the agent, not what the agent identity is authorized to call.
C is incorrect: Agents Toolkit is for scenarios such as custom SSO, middleware, multi-environment deployment, or advanced debugging, and it does not replace a missing role assignment.
D is incorrect: Portal publishing provisions Azure Bot Service as part of the Teams integration, so removing it breaks the channel rather than fixing permissions.
Q027 - Question
You saved a Foundry agent that uses the Azure Speech MCP server and Blob Storage for audio input and output. You need an application that lets users send speech requests to this agent outside the Foundry portal. What should you build?
Domain: Implement text analysis solutions (10–15%) Type: Single choice
- A. A Python client application that invokes the saved agent through the Foundry SDK.
- B. A Python client application that reads and writes Blob Storage directly and doesn't call the agent.
- C. A script that exports playground chat transcripts and replays them to users.
- D. A new agent for each user request, each with its own Azure Speech MCP server connection.
A is correct.
Explanation: Build a Python client application that invokes the saved agent through the Foundry SDK. The agent keeps its Azure Speech MCP server connection and selects the speech tools that each request needs.
B is incorrect: Blob Storage holds audio input and output, but it doesn't perform speech-to-text or text-to-speech. The application must invoke the agent.
C is incorrect: Replayed transcripts don't process new user requests. The application must invoke the saved agent at runtime.
D is incorrect: You connect the Azure Speech MCP server to the agent once and then invoke the saved agent. A new agent for each request adds setup work.
Q028 - Question
A developer saved a Foundry agent that is connected to the Azure Language MCP server and verified its behavior in the agent playground. The developer now needs a client application to call that same saved agent with the OpenAI Responses API. What should the developer do in the request?
Domain: Implement text analysis solutions (10–15%) Type: Single choice
- A. Supply the agent name in `agent_reference` through `extra_body`.
- B. Re-create the MCP tool definition in the client and send the tool list with every request.
- C. Pass the agent playground session identifier as the model name.
- D. Send the Azure Language endpoint URL as the response format parameter.
A is correct.
Explanation: A client application invokes the saved agent with the OpenAI Responses API by supplying the agent name in `agent_reference` through `extra_body`.
B is incorrect: The saved agent already holds the MCP tool connection, so the client does not need to resend a tool definition to reach the agent.
C is incorrect: The playground is used to test the agent, and its session identifier is not how the client references the saved agent.
D is incorrect: The response format parameter does not identify the agent, so the call would not reach the saved agent.
Q029 - Question
An ordering application must create a spoken version of each written order confirmation and save the audio as a file. Which approach should you use?
Domain: Implement text analysis solutions (10–15%) Type: Single choice
- A. Upload the confirmation text through the transcription operation to a gpt-4o-transcribe-diarize deployment
- B. Submit the confirmation text to a gpt-4o-mini-transcribe deployment and save the returned transcript
- C. Record the confirmation manually and upload the recording through the transcription operation
- D. Submit the confirmation text to a gpt-4o-mini-tts deployment and use the AzureOpenAI client to stream the generated speech to an output file
D is correct.
Explanation: Text-to-speech submits text to a model and returns an audio stream of vocalized text. gpt-4o-mini-tts is a supported synthesis model, and the AzureOpenAI client can stream the generated speech to an output file.
A is incorrect: gpt-4o-transcribe-diarize is a transcription model. The transcription operation takes audio input, not text for vocalization.
B is incorrect: gpt-4o-mini-transcribe converts speech to text. It returns a transcript, not an audio file.
C is incorrect: Uploading a recording through the transcription operation produces text. It does not generate speech from the confirmation text.
Q030 - Question
An organization maintains several voice client applications that use Voice Live. The team wants to centralize agent instructions and configuration in one place. The team also needs to support complex conversational logic while separating the agent logic from the voice clients. What should the team do?
Domain: Implement text analysis solutions (10–15%) Type: Single choice
- A. Configure the instructions in the WebSocket connection string for each client
- B. Embed the conversational logic directly in the Python client application
- C. Create a Foundry agent that contains the instructions and configuration, and connect the voice clients to the agent
- D. Store the conversational logic in the audio processing settings of each session
C is correct.
Explanation: Using a Foundry agent with Voice Live centralizes instructions and configuration, supports complex conversational logic, and separates agent logic from the voice client.
A is incorrect: Configuring instructions in the connection string for each client does not centralize the configuration or separate the logic effectively.
B is incorrect: Embedding logic directly in the client application fails to separate the agent logic from the voice client and requires updating each client.
D is incorrect: Audio processing settings are used for features like noise and echo reduction, not for storing conversational logic or agent instructions.
Q031 - Question
A marketing team wants to publish customer testimonials on a public website. Before publishing, the team must remove names, phone numbers, email addresses, and credit card numbers from the testimonial text. Which approach should you implement with Azure Language?
Domain: Implement text analysis solutions (10–15%) Type: Single choice
- A. Detect the language of each testimonial and publish only English submissions.
- B. Extract named entities and publish the original text with the entity list attached.
- C. Split each testimonial into documents under 5,120 characters before publishing.
- D. Use PII detection and return the redacted text for publication.
D is correct.
Explanation: Azure Language can identify sensitive personally identifiable information such as names, addresses, phone numbers, email addresses, social security numbers, and credit card numbers, and return redacted text to help protect privacy before the testimonials are published.
A is incorrect: Language detection returns a language identifier and score and leaves sensitive values in the published text.
B is incorrect: Publishing the original text still exposes the sensitive values, because entity extraction alone does not remove them.
C is incorrect: The 5,120 character limit applies to language detection requests and splitting text does not remove personal information.
Q032 - Question
Your speech agent uses the Azure Speech MCP server. It must transcribe call recordings stored in an Azure Blob Storage container. Company policy requires that the container remain private. How should you provide the audio to the agent?
Domain: Implement text analysis solutions (10–15%) Type: Single choice
- A. Give the agent a local file path to the recordings on a developer workstation.
- B. Add the storage account connection string to the user prompt.
- C. Provide a SAS URL for the audio, and treat the SAS URL as a secret.
- D. Change the container to allow public access so the agent can read the recordings.
C is correct.
Explanation: The Azure Speech MCP server transcribes audio from a public URL or a SAS URL. The container must stay private, so use a SAS URL. Treat SAS URLs as secrets because they grant access to the storage.
A is incorrect: Transcribed audio comes from a public URL or a SAS URL in Blob Storage, not from a local workstation path.
B is incorrect: The input for transcription is a public or SAS URL. Putting storage credentials in a prompt also exposes secret information.
D is incorrect: A public URL is a valid input, but public access violates the requirement to keep the container private.
Q033 - Question
A developer has a recorded meeting saved as an audio file. The application must return a text transcript by using a model deployed in Microsoft Foundry. What should the developer implement?
Domain: Implement text analysis solutions (10–15%) Type: Single choice
- A. Use an AzureOpenAI client to upload the audio file through the transcription operation to a gpt-4o-transcribe deployment
- B. Submit the audio file as input text to a gpt-4o-tts deployment
- C. Use an AzureOpenAI client to stream generated speech from the audio file to an output file
- D. Send a written summary of the meeting to a gpt-4o-mini-tts deployment
A is correct.
Explanation: Speech-to-text submits audio content to a model and returns a text transcript. gpt-4o-transcribe is a supported transcription model, and an AzureOpenAI client can upload an audio file through the transcription operation.
B is incorrect: gpt-4o-tts is a text-to-speech model. It takes text as input and returns audio, not a transcript.
C is incorrect: Streaming generated speech to an output file is part of speech synthesis. It does not create text from recorded audio.
D is incorrect: gpt-4o-mini-tts produces spoken audio from text. It cannot transcribe the original recording.
Q034 - Question
You're building a speech translation app with Azure Speech. A TranslationRecognizer returns text translations of spoken English into French, Spanish, and German. The app must also play spoken audio for all three translations. How should you produce the speech output?
Domain: Implement text analysis solutions (10–15%) Type: Single choice
- A. Use event-based synthesis to produce spoken output for all three target languages
- B. Use a SpeechSynthesizer for each translation
- C. Configure AudioConfig to return spoken output for each target language
- D. Use TextTranslationClient to convert each translation into speech
B is correct.
Explanation: For speech-to-speech output with multiple target languages, use a SpeechSynthesizer to synthesize each translation.
A is incorrect: Event-based synthesis applies to a single target language. This app has three target languages.
C is incorrect: AudioConfig specifies the audio input for translation. It doesn't produce the synthesized output.
D is incorrect: TextTranslationClient is used with Azure Translator for text translation. It doesn't synthesize speech.
Q035 - Question
A training company is implementing AI-powered translation with Microsoft Foundry. It has two requirements. Written chat messages from learners must be translated into other languages. Live spoken presentations must be translated for international attendees. Which two actions should you take? Each correct answer presents part of the solution.
Domain: Implement text analysis solutions (10–15%) Type: Multi-select
- A. Use the transliterate operation in Azure Translator to translate the spoken presentations
- B. Use Azure Translator in Foundry Tools to translate the written chat messages
- C. Use Azure Speech in Foundry Tools to translate the written chat messages
- D. Use Azure Speech in Foundry Tools to translate the spoken presentations
B and D are correct.
Explanation: Use Azure Translator in Foundry Tools for text translation, such as written chat messages. Use Azure Speech in Foundry Tools for speech translation, such as live spoken presentations.
A is incorrect: The transliterate operation converts text between scripts for a specified language. It doesn't translate spoken input.
C is incorrect: Azure Translator in Foundry Tools handles text translation. Azure Speech in Foundry Tools handles spoken input.
Q036 - Question
An architect reviews a design in which every text analysis tool signature is hardcoded into the agent definition. The team wants to reduce the maintenance work when the set of available tools changes. Which characteristic of the Azure Language MCP server supports this goal?
Domain: Implement text analysis solutions (10–15%) Type: Single choice
- A. The Model Context Protocol requires each tool to be registered as a separate model deployment.
- B. The server converts the agent prompt into a fixed sequence of Azure Language REST calls.
- C. The host, client, and server architecture supports runtime tool discovery, so the agent doesn't need hardcoded knowledge of each tool.
- D. The server stores the agent conversation history and replays it against each language tool.
C is correct.
Explanation: The Azure Language MCP server connects agents to Azure Language services through the Model Context Protocol. Its host, client, and server architecture supports runtime tool discovery, so the agent doesn't need hardcoded knowledge of each tool.
A is incorrect: The tools are exposed by the MCP server, not registered as individual model deployments.
B is incorrect: The agent selects tools at runtime rather than running a fixed call sequence defined by the server.
D is incorrect: Conversation replay is not a described function of the MCP server architecture.
Q037 - Question
A support portal receives customer messages in many languages. The application doesn't know the language of each message in advance. All messages must be translated into English before agents review them. You use TextTranslationClient with Azure Translator. What should you do?
Domain: Implement text analysis solutions (10–15%) Type: Single choice
- A. Call the transliterate operation and specify English as the target script
- B. Call the translate operation and always specify Spanish as the source language
- C. Call the transliterate operation first to identify the source language, and then call translate
- D. Call the translate operation with English as the target language and omit the source language
D is correct.
Explanation: The translate operation can detect the source language when you omit it. Specify English as the target language and let the service identify the language of each message.
A is incorrect: The transliterate operation converts text between scripts for a specified language. It doesn't translate text into another language.
B is incorrect: A fixed source language doesn't match messages that arrive in many unknown languages.
C is incorrect: The transliterate operation requires a specified language, so you can't use it to identify an unknown source language.
Q038 - Question
A retail company plans to build an AI agent that transcribes customer voicemails and generates spoken replies. Users make requests in natural language. The developers want to avoid code that routes each request to a specific speech operation. What should you do?
Domain: Implement text analysis solutions (10–15%) Type: Single choice
- A. Create one agent for speech-to-text and another for text-to-speech, and write routing logic that selects an agent for each request.
- B. Connect the agent to the Azure Speech MCP server through a single tool connection.
- C. Write custom function code that parses each request and calls the matching speech operation.
- D. Configure the agent for text-to-speech only, and handle transcription in separate client code.
B is correct.
Explanation: The Azure Speech MCP server makes speech-to-text and text-to-speech available to an agent through a single tool connection. The agent handles speech tasks based on the user's natural language request, so you don't need operation-specific routing code.
A is incorrect: Separate agents with routing logic add the operation-specific routing code that the requirement is meant to avoid.
C is incorrect: Custom parsing and dispatch code is operation-specific routing. The Azure Speech MCP server lets the agent handle the request without this code.
D is incorrect: This approach splits the speech capabilities across components. The Azure Speech MCP server provides both capabilities to the agent through one connection.
Q039 - Question
A developer must build a proof of concept that extracts structured data from product images by using a custom schema. The developer plans to call the Content Understanding API from a Python application. What should the developer do first, and what is the prerequisite?
Domain: Implement computer vision solutions (10–15%) Type: Single choice
- A. Create a custom image analyzer in the Microsoft Foundry portal; an Azure subscription is required
- B. Write the Python application first and define the schema inline at run time; no Azure subscription is required
- C. Publish the images to a public website so the API can read them; an Azure subscription is required
- D. Convert each image to Markdown locally before calling the API; no Azure subscription is required
A is correct.
Explanation: You must create a custom image analyzer in the Microsoft Foundry portal first, and then create the Python app that calls the Content Understanding API to analyze images. You need an Azure subscription to complete this process.
B is incorrect: The application submits an analysis request by analyzer ID, so the analyzer must exist first in the portal, and an Azure subscription is required.
C is incorrect: Publishing images to a public website is not required to create the analyzer, and it does not replace the step of creating the analyzer in the portal.
D is incorrect: Markdown is part of the structured analysis result returned by the service, not an input you create locally, and an Azure subscription is required.
Q040 - Question
A finance team wants Content Understanding to return specific values, such as supplier name and total amount, from purchase order documents. The team has already created a Foundry resource. What should the team do next?
Domain: Implement computer vision solutions (10–15%) Type: Single choice
- A. Define a schema for the information to extract, based on a sample document and a template.
- B. Submit production purchase orders to the analyzer-results endpoint.
- C. Publish a new version of the analyzer before any fields are defined.
- D. Convert every purchase order to an image file before any configuration.
A is correct.
Explanation: An analyzer is built from a schema that defines the information to extract from a content type, and you define that schema from a sample and a template before you build the analyzer.
B is incorrect: You can only analyze new content after the analyzer is built from a schema.
C is incorrect: Versioning applies to an analyzer that already has a defined schema.
D is incorrect: Content Understanding analyzes documents as well as images, so converting the files is not a required step.
Q041 - Question
A company stores contracts as PDF files, product data in a relational database, and support notes as free text. The team requires one solution that can index all three content types, enrich the indexed content with AI, and supply grounding data for a retrieval augmented generation (RAG) application. Which capability of Azure AI Search supports using a single platform for this workload?
Domain: Implement information extraction solutions (10–15%) Type: Single choice
- A. Azure AI Search indexes structured, semi-structured, and unstructured data, enriches data with AI skills, and supports vector-based RAG grounding.
- B. Azure AI Search indexes structured and unstructured data natively but requires an external service to enrich the data with AI skills.
- C. Azure AI Search enriches unstructured data with AI skills but cannot index structured relational data.
- D. Azure AI Search supports enterprise search and knowledge mining but relies on an external service for vector-based RAG grounding.
A is correct.
Explanation: Azure AI Search indexes structured, semi-structured, and unstructured data, can enrich that data with AI skills, and supports enterprise search, knowledge mining, and vector-based grounding for RAG scenarios.
B is incorrect: Azure AI Search does not require an external service for enrichment; it uses an internal enrichment pipeline with AI skills.
C is incorrect: Azure AI Search supports indexing structured data in addition to semi-structured and unstructured data.
D is incorrect: Azure AI Search natively supports vector-based RAG grounding and does not rely on an external service for this capability.
Q042 - Question
A developer uses the OpenAI Python SDK to create Sora 2 video jobs that start from a reference image. Several jobs are rejected. Which two actions should the developer take to make the reference images usable? Each correct answer presents part of the solution.
Domain: Implement computer vision solutions (10–15%) Type: Multiple choice
- A. Convert each reference image to grayscale before submitting the job.
- B. Resize each reference image so its resolution exactly matches the output video size.
- C. Increase the video duration so the model has more time to process the image.
- D. Replace reference images that contain human faces with images that do not.
B and D are correct.
Explanation: A reference image must have a resolution that exactly matches the requested output video size, and reference images containing human faces are rejected.
A is incorrect: Color is not a constraint on reference images; the resolution match and the restriction on human faces determine acceptance.
C is incorrect: Duration sets the length of the generated video and does not affect whether a reference image is accepted.
Q043 - Question
A team uses Azure AI Search to index scanned agreements. The solution must extract text from the scanned page images, detect key phrases in that text, and extract custom structured fields that no built-in skill produces. Which two design actions should you include in the enrichment pipeline? Each correct answer presents part of the solution.
Domain: Implement information extraction solutions (10–15%) Type: Multiple choice
- A. Query the index using full Lucene syntax to extract text at query time.
- B. Configure built-in AI skills that use Foundry Tools capabilities for OCR and key phrase detection.
- C. Configure a table projection in a knowledge store to generate the missing structured fields.
- D. Configure a custom skill that wraps a service such as Document Intelligence to produce the missing structured fields.
B and D are correct.
Explanation: Built-in skills use Foundry Tools capabilities for tasks such as key-phrase detection and OCR. When no built-in skill produces the necessary fields, a custom skill can wrap services such as Document Intelligence to extract the custom structured fields.
A is incorrect: Queries evaluate an existing index through parsing and retrieval; they do not perform OCR extraction from images.
C is incorrect: A table projection persists enriched data in a relational schema for reporting; it does not generate new enriched fields during the pipeline process.
Q044 - Question
You are developing a vision-based chat app that uses the Responses API. Some images are available at a web address. Other images are stored locally on the user's device. Which two methods can you use to include images in the prompt? Each correct answer presents a complete solution.
Domain: Implement computer vision solutions (10–15%) Type: Multi-select
- A. Include the local file path of the image as plain text in the user message
- B. Provide the image URL in the image content of the user message
- C. Provide Base64-encoded local image data in a data URL
- D. Include only the image file name in the text content of the user message
B and C are correct.
Explanation: Responses API and ChatCompletions API inputs can use an image URL. For local images, you can use Base64-encoded image data in a data URL.
A is incorrect: A local file path is text that points to a location on the user's device. For local images, encode the image data as Base64 in a data URL.
D is incorrect: A file name is text and doesn't provide the image content. Use an image URL or a Base64 data URL instead.
Q045 - Question
A manufacturer builds a custom image analyzer for inspection photos. One field must capture the serial number printed on the part. Another field must assign each photo to one of four defined damage categories. Which two extraction methods should the team configure for these fields? (Choose two.)
Domain: Implement computer vision solutions (10–15%) Type: Multiple choice
- A. Extract for the serial number field
- B. Generate for the serial number field
- C. Classify for the damage category field
- D. Classify for the serial number field
A and C are correct.
Explanation: Each field in a custom image schema can use the extract, classify, or generate method. Use extract to return a value that is visible in the image, such as a printed serial number. Use classify to assign the image to one of a defined set of options, such as damage categories.
B is incorrect: Generate produces a description rather than returning the exact printed value, so it is not the correct method for extracting a visible serial number.
D is incorrect: Classify chooses from a fixed set of predefined options, so it cannot return an arbitrary extracted serial number value.
Q046 - Question
A team must create Content Understanding analyzers automatically as part of a deployment pipeline. The pipeline uses REST calls to interact with the API. Which approach should the team configure the pipeline to use?
Domain: Implement computer vision solutions (10–15%) Type: Single choice
- A. Export a trained analyzer from Content Understanding Studio at run time and import it with a GET request.
- B. Submit a JSON analyzer schema with a PUT request, then poll the URL in the Operation-Location response.
- C. Send a content file to the endpoint first and let the service generate an analyzer from the detected fields.
- D. Create the analyzer with a POST request that returns the finished analyzer definition in the synchronous response.
B is correct.
Explanation: With REST, you submit a JSON analyzer schema using PUT and then poll the Operation-Location value returned in the response, because analyzer creation is asynchronous.
A is incorrect: Exporting a trained analyzer and importing it with a GET request does not programmatically create an analyzer in a pipeline.
C is incorrect: Submitting a content file to the service analyzes the content but does not generate an analyzer schema.
D is incorrect: Analyzer creation is asynchronous and does not complete within a single synchronous POST request response.
Q047 - Question
A media company configures a client application to process contracts, product photos, call recordings, and training videos. The application writes the extracted values into a structured database. Which statement correctly describes the output the application receives from Content Understanding?
Domain: Implement computer vision solutions (10–15%) Type: Single choice
- A. Each content type requires a separate service deployment, because a single analyzer API cannot accept audio or video data.
- B. Results are returned only as plain text, so the application must parse structured field values from the text block.
- C. The application receives typed field values, Markdown, and metadata based on the analyzer schema and the content.
- D. The application receives the original file with annotations, and field extraction must be added by a separate model.
C is correct.
Explanation: Analysis outputs expose typed fields, Markdown, and metadata based on the analyzer schema and the content, allowing applications to write type-specific values directly into a structured database.
A is incorrect: You can submit binary document, image, audio, or video data to a single analyzer API.
B is incorrect: Extracted fields are returned as typed values, not just plain unstructured text.
D is incorrect: The application receives the extracted or generated field values from the analyzer, so no separate extraction model is required.
Q048 - Question
You are designing an Azure AI Search solution. The solution must read documents from a data source, run an enrichment pipeline over the extracted content, and place both source fields and skill output fields into an index. Which component should you configure to perform these tasks?
Domain: Implement information extraction solutions (10–15%) Type: Single choice
- A. A knowledge store that writes projections to storage.
- B. A search query that uses full Lucene syntax.
- C. An indexer that connects to the data source.
- D. An OData filter expression.
C is correct.
Explanation: An Azure AI Search indexer connects to a data source, applies document cracking and an enrichment pipeline, and maps source and skill output fields to an index.
A is incorrect: A knowledge store persists projections of enriched data to storage; it does not read from the data source or populate the index.
B is incorrect: A search query evaluates search expressions against an existing index; it does not extract or enrich data.
D is incorrect: An OData filter expression evaluates criteria to filter query results; it is not used to extract or enrich source data.
Q049 - Question
A development team plans to integrate Azure Document Intelligence into a claims processing application. The team needs the analysis results in a machine-readable format that includes the position of each extracted element on the page, so the application can highlight the source region for every value it stores. Which characteristic of the service output meets this requirement?
Domain: Implement information extraction solutions (10–15%) Type: Single choice
- A. Structured JSON that includes text, key-value pairs, tables, and bounding-box data
- B. A plain text transcript of the document with no positional information
- C. A rendered image file with the extracted values drawn onto the page
- D. A relational database table that the application queries directly
A is correct.
Explanation: Azure Document Intelligence returns structured JSON that contains text, key-value pairs, selection marks, tables, and bounding-box data, so your application can map each extracted value back to its position on the page.
B is incorrect: A plain text transcript omits the bounding-box data needed to highlight source regions.
C is incorrect: The output is structured JSON, not a rendered image with values drawn on the page.
D is incorrect: The service returns JSON through REST, SDKs, Studio, or Foundry, not a relational database table.
Q050 - Question
A design team wants to test prompt wording and review generated images before any application code is written. Which action meets this requirement with the least development effort?
Domain: Implement computer vision solutions (10–15%) Type: Single choice
- A. Write a console application that calls the Images API for each prompt variation.
- B. Deploy a container image of the model to a virtual machine and post prompts to it.
- C. Create a Foundry project and submit prompts in the model playground in Microsoft Foundry portal.
- D. Export the model weights and run prompt tests on a local workstation.
C is correct.
Explanation: After you create a Foundry project, you can use the model playground in Microsoft Foundry portal to submit prompts and view the resulting generated images, and when supported you can set the generated-image resolution and include a reference image to guide output.
A is incorrect: Writing client code is a valid path for applications, but it adds development work that the playground avoids for prompt testing.
B is incorrect: The described way to test prompts interactively is the portal playground, not self-hosting on a virtual machine.
D is incorrect: Prompt testing is done through the portal playground or a client that calls the model, not by exporting model weights.