Microsoft AI-102 (Designing and Implementing a Microsoft Azure AI Solution) Practice Test 20

Question 190

An app uses developer credentials locally but must run unattended in Azure production. What authentication change is preferred?

  1. Use a deployed workload identity with supported role assignments instead of interactive login
  2. Create the named deployment and use its exact identifier in inference requests
  3. Use a secured project connection or approved secret injection for the external call
  4. Examine failure examples, subgroup performance, and safety results alongside aggregates

Explanation

Correct Answer

A:

Production automation should authenticate as a governed principal rather than relying on a developer's session. In the team-developed knowledge assistant, confirm the fix by retesting the affected behavior after using a deployed workload identity with supported role assignments instead of interactive login.

Incorrect Answers

B:

Although useful elsewhere in the team-developed knowledge assistant, catalog listings describe available models, while inference addresses actual deployed instances; the present case is different and calls for the team to use a deployed workload identity with supported role assignments instead of interactive login.

C:

That change addresses another failure mode: flows should resolve managed credentials; embedding secrets in workflow definitions risks exposure; it does not address this case, for which the relevant intervention is to use a deployed workload identity with supported role assignments instead of interactive login.

D:

Mean scores can conceal severe errors on rare or high-impact cases in its own workflow, but that is a different correction from the present requirement to use a deployed workload identity with supported role assignments instead of interactive login. The reported issue would remain.

Question 191

A developer assumes creating a Foundry project automatically provisions a model endpoint. What action is still necessary?

  1. Freeze and version the evaluation dataset and scoring criteria for regression runs
  2. Create a compatible model deployment and address it from the application
  3. Pin compatible SDK dependency versions and validate upgrades in staging
  4. Parameterize the prompt-flow inputs and run both variants on a shared test dataset

Explanation

Correct Answer

B:

A project organizes assets; it is not itself an inference deployment. In the team-developed knowledge assistant, confirm the fix by retesting the affected behavior after creating a compatible model deployment and address it from the application.

Incorrect Answers

A:

Although useful elsewhere in the team-developed knowledge assistant, controlled datasets and criteria allow meaningful before/after comparisons; the present case is different and calls for the team to create a compatible model deployment and address it from the application.

C:

Reproducible dependency versions prevent silent breaking changes during releases in its own workflow, but that is a different correction from the present requirement to create a compatible model deployment and address it from the application. The reported issue would remain.

D:

That change addresses another failure mode: repeatable parameterized flow runs support controlled comparisons rather than ad hoc edits; it does not address this case, for which the relevant intervention is to create a compatible model deployment and address it from the application.

Question 192

Two applications in the same subscription use unrelated models and indexes. How should evaluation assets be organized?

  1. Use a deployed workload identity with supported role assignments instead of interactive login
  2. Create the named deployment and use its exact identifier in inference requests
  3. Examine failure examples, subgroup performance, and safety results alongside aggregates
  4. Associate each application's evaluation datasets and runs with its intended project

Explanation

Correct Answer

D:

Separating project assets avoids mixing comparisons and experiment history across products. In the team-developed knowledge assistant, confirm the fix by retesting the affected behavior after associating each application's evaluation datasets and runs with its intended project.

Incorrect Answers

A:

That change addresses another failure mode: production automation should authenticate as a governed principal rather than relying on a developer's session; it does not address this case, for which the relevant intervention is to associate each application's evaluation datasets and runs with its intended project.

B:

Although useful elsewhere in the team-developed knowledge assistant, catalog listings describe available models, while inference addresses actual deployed instances; the present case is different and calls for the team to associate each application's evaluation datasets and runs with its intended project.

C:

Mean scores can conceal severe errors on rare or high-impact cases in its own workflow, but that is a different correction from the present requirement to associate each application's evaluation datasets and runs with its intended project. The reported issue would remain.

Question 193

A new evaluator cannot reproduce a teammate's experiments because connections live in personal local files. What workspace practice helps?

  1. Score both response quality and groundedness on the same held-out examples
  2. Use governed project connections and shared evaluation assets rather than private local settings
  3. Load both the intended Foundry project connection and the target model deployment
  4. Use a model deployment documented to support the requested image-input modality

Explanation

Correct Answer

B:

Project-managed artifacts let collaborators use a consistent application environment while retaining access controls. In the team-developed knowledge assistant, confirm the fix by retesting the affected behavior after using governed project connections and shared evaluation assets rather than private local settings.

Incorrect Answers

A:

Although useful elsewhere in the team-developed knowledge assistant, single-metric fluency assessments miss whether a response faithfully uses retrieved evidence; the present case is different and calls for the team to use governed project connections and shared evaluation assets rather than private local settings.

C:

That change addresses another failure mode: sDK clients still need concrete inference targets; a project endpoint alone does not name the model to serve a request; it does not address this case, for which the relevant intervention is to use governed project connections and shared evaluation assets rather than private local settings.

D:

Endpoint availability does not imply every model accepts every input type in its own workflow, but that is a different correction from the present requirement to use governed project connections and shared evaluation assets rather than private local settings. The reported issue would remain.

Question 194

A flow calls an external service with a private credential hard-coded into a node. How should it be refactored?

  1. Examine failure examples, subgroup performance, and safety results alongside aggregates
  2. Associate each application's evaluation datasets and runs with its intended project
  3. Create the named deployment and use its exact identifier in inference requests
  4. Use a secured project connection or approved secret injection for the external call

Explanation

Correct Answer

D:

Flows should resolve managed credentials; embedding secrets in workflow definitions risks exposure. In the team-developed knowledge assistant, confirm the fix by retesting the affected behavior after using a secured project connection or approved secret injection for the external call.

Incorrect Answers

A:

Mean scores can conceal severe errors on rare or high-impact cases in its own workflow, but that is a different correction from the present requirement to use a secured project connection or approved secret injection for the external call. The reported issue would remain.

B:

Although useful elsewhere in the team-developed knowledge assistant, separating project assets avoids mixing comparisons and experiment history across products; the present case is different and calls for the team to use a secured project connection or approved secret injection for the external call.

C:

That change addresses another failure mode: catalog listings describe available models, while inference addresses actual deployed instances; it does not address this case, for which the relevant intervention is to use a secured project connection or approved secret injection for the external call.

Question 195

A model is visible in the catalog but the request references a nickname not deployed in the project. What is required?

  1. Associate each application's evaluation datasets and runs with its intended project
  2. Examine failure examples, subgroup performance, and safety results alongside aggregates
  3. Create the named deployment and use its exact identifier in inference requests
  4. Use a deployed workload identity with supported role assignments instead of interactive login

Explanation

Correct Answer

C:

Catalog listings describe available models, while inference addresses actual deployed instances. In the team-developed knowledge assistant, confirm the fix by retesting the affected behavior after creating the named deployment and use its exact identifier in inference requests.

Incorrect Answers

A:

Although useful elsewhere in the team-developed knowledge assistant, separating project assets avoids mixing comparisons and experiment history across products; the present case is different and calls for the team to create the named deployment and use its exact identifier in inference requests.

B:

That change addresses another failure mode: mean scores can conceal severe errors on rare or high-impact cases; it does not address this case, for which the relevant intervention is to create the named deployment and use its exact identifier in inference requests.

D:

Production automation should authenticate as a governed principal rather than relying on a developer's session in its own workflow, but that is a different correction from the present requirement to create the named deployment and use its exact identifier in inference requests. The reported issue would remain.

Question 196

The solution brief for a team-developed knowledge assistant requires the system to compare response quality and safety on representative inputs against specified criteria. A successful test confirms that evaluation results reveal regressions across a fixed test dataset. What should the team use?

  1. Build a prompt flow
  2. Deploy an appropriate model from the catalog
  3. Integrate the application through the Foundry SDK
  4. Use a Foundry project workspace
  5. Run a model or flow evaluation

Explanation

Correct Answer

E:

Evaluation applies metrics or evaluators to test responses, enabling systematic comparison of changes. For this team-developed knowledge assistant, the intended evidence is that evaluation results reveal regressions across a fixed test dataset. It is distinct from publishing a model endpoint or buying throughput capacity.

Incorrect Answers

A:

Prompt flow represents a connected sequence of executable steps that can be tested and compared; it may be used to connect prompt, model, and processing nodes in a reproducible workflow in a team-developed knowledge assistant, but the current requirement is instead to compare response quality and safety on representative inputs against specified criteria. That separate capability does not prove that evaluation results reveal regressions across a fixed test dataset; It does not automatically make all enterprise documents available to a model, so selecting that alternative would not meet the stated acceptance check.

B:

A model deployment provides an inference target with a deployment name and the selected model's capabilities; it may be used to make a selected compatible generative model callable by application code in a team-developed knowledge assistant, but the current requirement is instead to compare response quality and safety on representative inputs against specified criteria. That separate capability does not prove that evaluation results reveal regressions across a fixed test dataset; Selecting a catalog entry alone does not produce a callable deployment, so selecting that alternative would not meet the stated acceptance check.

C:

The SDK provides supported programmatic interfaces for the selected Foundry environment and API surface; it may be used to call project resources and supported model or agent operations in application code in a team-developed knowledge assistant, but the current requirement is instead to compare response quality and safety on representative inputs against specified criteria. That separate capability does not prove that evaluation results reveal regressions across a fixed test dataset; Installing an SDK does not eliminate the need for authentication and deployed resources, so selecting that alternative would not meet the stated acceptance check.

D:

A project provides the working context for developing AI applications and managing associated resources; it may be used to keep the application's resources, connections, and evaluation assets organized for developers in a team-developed knowledge assistant, but the current requirement is instead to compare response quality and safety on representative inputs against specified criteria. That separate capability does not prove that evaluation results reveal regressions across a fixed test dataset; The project itself is not a text embedding model or search query, so selecting that alternative would not meet the stated acceptance check.

Question 197

In a team-developed knowledge assistant, engineers must make a selected compatible generative model callable by application code. The expected evidence is that a test client can invoke the named deployed model. Which implementation is the best match?

  1. Use a Foundry project workspace
  2. Deploy an appropriate model from the catalog
  3. Build a prompt flow
  4. Run a model or flow evaluation

Explanation

Correct Answer

B:

A model deployment provides an inference target with a deployment name and the selected model's capabilities. For this team-developed knowledge assistant, the intended evidence is that a test client can invoke the named deployed model. Selecting a catalog entry alone does not produce a callable deployment.

Incorrect Answers

A:

A project provides the working context for developing AI applications and managing associated resources; it may be used to keep the application's resources, connections, and evaluation assets organized for developers in a team-developed knowledge assistant, but the current requirement is instead to make a selected compatible generative model callable by application code. That separate capability does not prove that a test client can invoke the named deployed model; The project itself is not a text embedding model or search query, so selecting that alternative would not meet the stated acceptance check.

C:

Prompt flow represents a connected sequence of executable steps that can be tested and compared; it may be used to connect prompt, model, and processing nodes in a reproducible workflow in a team-developed knowledge assistant, but the current requirement is instead to make a selected compatible generative model callable by application code. That separate capability does not prove that a test client can invoke the named deployed model; It does not automatically make all enterprise documents available to a model, so selecting that alternative would not meet the stated acceptance check.

D:

Evaluation applies metrics or evaluators to test responses, enabling systematic comparison of changes; it may be used to compare response quality and safety on representative inputs against specified criteria in a team-developed knowledge assistant, but the current requirement is instead to make a selected compatible generative model callable by application code. That separate capability does not prove that a test client can invoke the named deployed model; It is distinct from publishing a model endpoint or buying throughput capacity, so selecting that alternative would not meet the stated acceptance check.

Question 198

A Python app accesses a project but lacks a model deployment identifier. What should its configuration resolve explicitly?

  1. Load both the intended Foundry project connection and the target model deployment
  2. Use governed project connections and shared evaluation assets rather than private local settings
  3. Use a model deployment documented to support the requested image-input modality
  4. Capture intermediate node inputs and outputs for the flow run

Explanation

Correct Answer

A:

SDK clients still need concrete inference targets; a project endpoint alone does not name the model to serve a request. In the team-developed knowledge assistant, confirm the fix by retesting the affected behavior after loading both the intended Foundry project connection and the target model deployment.

Incorrect Answers

B:

That change addresses another failure mode: project-managed artifacts let collaborators use a consistent application environment while retaining access controls; it does not address this case, for which the relevant intervention is to load both the intended Foundry project connection and the target model deployment.

C:

Although useful elsewhere in the team-developed knowledge assistant, endpoint availability does not imply every model accepts every input type; the present case is different and calls for the team to load both the intended Foundry project connection and the target model deployment.

D:

Node-level execution results make data movement and failures observable across multi-step prompt processing in its own workflow, but that is a different correction from the present requirement to load both the intended Foundry project connection and the target model deployment. The reported issue would remain.

Question 199

An agent benchmark reports different scores on every run because the input samples change. What practice makes comparisons fair?

  1. Pin compatible SDK dependency versions and validate upgrades in staging
  2. Freeze and version the evaluation dataset and scoring criteria for regression runs
  3. Target the intended production deployment name when running release validation
  4. Create a compatible model deployment and address it from the application

Explanation

Correct Answer

B:

Controlled datasets and criteria allow meaningful before/after comparisons. In the team-developed knowledge assistant, confirm the fix by retesting the affected behavior after freezing and version the evaluation dataset and scoring criteria for regression runs.

Incorrect Answers

A:

Although useful elsewhere in the team-developed knowledge assistant, reproducible dependency versions prevent silent breaking changes during releases; the present case is different and calls for the team to freeze and version the evaluation dataset and scoring criteria for regression runs.

C:

That change addresses another failure mode: test results are meaningful only for the deployment clients will actually call; it does not address this case, for which the relevant intervention is to freeze and version the evaluation dataset and scoring criteria for regression runs.

D:

A project organizes assets; it is not itself an inference deployment in its own workflow, but that is a different correction from the present requirement to freeze and version the evaluation dataset and scoring criteria for regression runs. The reported issue would remain.

Question 200

Production changed model versions but tests still invoke staging. What should be corrected in the client?

  1. Create a compatible model deployment and address it from the application
  2. Freeze and version the evaluation dataset and scoring criteria for regression runs
  3. Target the intended production deployment name when running release validation
  4. Pin compatible SDK dependency versions and validate upgrades in staging

Explanation

Correct Answer

C:

Test results are meaningful only for the deployment clients will actually call. In the team-developed knowledge assistant, confirm the fix by retesting the affected behavior after targeting the intended production deployment name when running release validation.

Incorrect Answers

A:

That change addresses another failure mode: a project organizes assets; it is not itself an inference deployment; it does not address this case, for which the relevant intervention is to target the intended production deployment name when running release validation.

B:

Controlled datasets and criteria allow meaningful before/after comparisons in its own workflow, but that is a different correction from the present requirement to target the intended production deployment name when running release validation. The reported issue would remain.

D:

Although useful elsewhere in the team-developed knowledge assistant, reproducible dependency versions prevent silent breaking changes during releases; the present case is different and calls for the team to target the intended production deployment name when running release validation.

Designing and Implementing a Microsoft Azure AI Solution (AI-102) Practice Tests