Microsoft AI-103 (Developing AI Apps and Agents on Azure) Practice Test 4

Question 30

A new customer-facing AI service must retrieve entries containing exact product part numbers and uncommon policy codes. Which approach should the team select first?

  1. Ground responses in a search knowledge source
  2. Apply semantic reranking to candidate passages
  3. Choose full-text keyword search for precise identifiers
  4. Use vector similarity search for paraphrases

Explanation

Correct Answer

C:

For the stated requirement to retrieve entries containing exact product part numbers and uncommon policy codes, keyword retrieval matches specific lexical tokens precisely, important when identifiers lack semantic meaning to an embedding model; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to answer policy questions using the latest approved internal manuals instead of model recollection, a connected retrieval source gives an agent current authorized reference content that can be cited and updated separately from model weights; this does not directly meet the required outcome to retrieve entries containing exact product part numbers and uncommon policy codes.

B:

For a system whose objective is to improve ordering of an already retrieved set by how directly it answers a user's question, semantic reranking uses language understanding to reorder candidate search results; it does not replace indexing or generate new source documents; this does not directly meet the required outcome to retrieve entries containing exact product part numbers and uncommon policy codes.

D:

For a system whose objective is to find related guidance when employees use very different wording from the source documents, vector search compares embeddings to retrieve semantically close text even when terms differ, provided index and query embeddings are compatible; this does not directly meet the required outcome to retrieve entries containing exact product part numbers and uncommon policy codes.

Question 31

The team plans to choose full-text keyword search for precise identifiers. Which test provides the clearest evidence that this change works?

  1. Track normalized discounted cumulative gain and top-rank relevance on judged query results
  2. Measure exact-identifier match recall and precision on known part-number queries
  3. Compare retrieval success on mixed identifier-and-paraphrase evaluation queries
  4. Audit answer support against current source passages and check citation correctness

Explanation

Correct Answer

B:

The accepted change is to choose full-text keyword search for precise identifiers; this evaluation checks whether it can retrieve entries containing exact product part numbers and uncommon policy codes, since keyword retrieval matches specific lexical tokens precisely, important when identifiers lack semantic meaning to an embedding model.

Incorrect Answers

A:

This test concerns the ability to improve ordering of an already retrieved set by how directly it answers a user's question, for which semantic reranking uses language understanding to reorder candidate search results; it does not replace indexing or generate new source documents; it does not establish whether the change to choose full-text keyword search for precise identifiers works for the current release.

C:

This test concerns the ability to support queries mixing specific document codes with natural-language intent, for which hybrid retrieval runs lexical and vector queries together and merges ranked results; it helps preserve exact tokens while matching semantic concepts; it does not establish whether the change to choose full-text keyword search for precise identifiers works for the current release.

D:

This test concerns the ability to answer policy questions using the latest approved internal manuals instead of model recollection, for which a connected retrieval source gives an agent current authorized reference content that can be cited and updated separately from model weights; it does not establish whether the change to choose full-text keyword search for precise identifiers works for the current release.

Question 32

A solution architect is deciding how to categorize short support tickets under strict response-time and inference-budget limits. Which Azure AI pattern addresses that task?

  1. Use an embedding model to represent meaning
  2. Use vector similarity search for paraphrases
  3. Deploy a multimodal model for visual reasoning
  4. Select a small language model for bounded classification

Explanation

Correct Answer

D:

For the stated requirement to categorize short support tickets under strict response-time and inference-budget limits, a compact model can meet narrow classification needs with fewer inference resources when validated against representative labeled examples; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to map queries and reference passages into vectors for semantic retrieval, embeddings encode text into vectors used by a similarity search system; they are not themselves a response-generating chatbot; this does not directly meet the required outcome to categorize short support tickets under strict response-time and inference-budget limits.

B:

For a system whose objective is to find related guidance when employees use very different wording from the source documents, vector search compares embeddings to retrieve semantically close text even when terms differ, provided index and query embeddings are compatible; this does not directly meet the required outcome to categorize short support tickets under strict response-time and inference-budget limits.

C:

For a system whose objective is to answer natural-language questions about photographed equipment where the answer depends on visible details, a multimodal model accepts image evidence alongside text; a text-only model cannot inspect visual input unless a separate vision pipeline supplies it; this does not directly meet the required outcome to categorize short support tickets under strict response-time and inference-budget limits.

Question 33

A pilot has a functional gap: Questions containing a form number and a descriptive phrase are poorly served by either lexical-only or vector-only retrieval. How should the team address that gap?

  1. Persist only allowed preference fields behind identity checks and expiration controls
  2. Submit combined text and vector queries to a hybrid-capable search index
  3. Pass the actual image to a vision-capable model and ground its response in visible evidence
  4. Run vector retrieval on document chunks using the same embedding space as queries
  5. Enable semantic ranking over searchable human-readable fields after initial retrieval

Explanation

Correct Answer

B:

The fault signal is that questions containing a form number and a descriptive phrase are poorly served by either lexical-only or vector-only retrieval; the targeted fix works because hybrid retrieval runs lexical and vector queries together and merges ranked results; it helps preserve exact tokens while matching semantic concepts, restoring the ability to support queries mixing specific document codes with natural-language intent.

Incorrect Answers

A:

If the failing function were to personalize repeated interactions without placing all old transcripts in every prompt, this change could be relevant because explicitly persisted, consent-governed preference data supports bounded long-term personalization; transient chat state should not become an uncontrolled permanent profile; the actual symptom is that questions containing a form number and a descriptive phrase are poorly served by either lexical-only or vector-only retrieval, which calls for a different repair.

C:

If the failing function were to answer natural-language questions about photographed equipment where the answer depends on visible details, this change could be relevant because a multimodal model accepts image evidence alongside text; a text-only model cannot inspect visual input unless a separate vision pipeline supplies it; the actual symptom is that questions containing a form number and a descriptive phrase are poorly served by either lexical-only or vector-only retrieval, which calls for a different repair.

D:

If the failing function were to find related guidance when employees use very different wording from the source documents, this change could be relevant because vector search compares embeddings to retrieve semantically close text even when terms differ, provided index and query embeddings are compatible; the actual symptom is that questions containing a form number and a descriptive phrase are poorly served by either lexical-only or vector-only retrieval, which calls for a different repair.

E:

If the failing function were to improve ordering of an already retrieved set by how directly it answers a user's question, this change could be relevant because semantic reranking uses language understanding to reorder candidate search results; it does not replace indexing or generate new source documents; the actual symptom is that questions containing a form number and a descriptive phrase are poorly served by either lexical-only or vector-only retrieval, which calls for a different repair.

Question 34

A Foundry-based application is being scoped for production. The non-negotiable task is to find related guidance when employees use very different wording from the source documents. Which implementation fits?

  1. Combine keyword and vectors in hybrid search
  2. Use vector similarity search for paraphrases
  3. Use an embedding model to represent meaning
  4. Choose full-text keyword search for precise identifiers

Explanation

Correct Answer

B:

For the stated requirement to find related guidance when employees use very different wording from the source documents, vector search compares embeddings to retrieve semantically close text even when terms differ, provided index and query embeddings are compatible; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to support queries mixing specific document codes with natural-language intent, hybrid retrieval runs lexical and vector queries together and merges ranked results; it helps preserve exact tokens while matching semantic concepts; this does not directly meet the required outcome to find related guidance when employees use very different wording from the source documents.

C:

For a system whose objective is to map queries and reference passages into vectors for semantic retrieval, embeddings encode text into vectors used by a similarity search system; they are not themselves a response-generating chatbot; this does not directly meet the required outcome to find related guidance when employees use very different wording from the source documents.

D:

For a system whose objective is to retrieve entries containing exact product part numbers and uncommon policy codes, keyword retrieval matches specific lexical tokens precisely, important when identifiers lack semantic meaning to an embedding model; this does not directly meet the required outcome to find related guidance when employees use very different wording from the source documents.

Question 35

Production telemetry reveals this failure: The assistant summarizes historical order records but cannot report a shipment that changed status minutes ago. Which change most directly corrects it?

  1. Persist only allowed preference fields behind identity checks and expiration controls
  2. Embed the queries and chunks with a compatible model and add vector retrieval
  3. Connect governed document retrieval and pass cited passages to the answer step
  4. Register an approved order-status API tool and scope its credentials to read-only lookup

Explanation

Correct Answer

D:

The fault signal is that the assistant summarizes historical order records but cannot report a shipment that changed status minutes ago; the targeted fix works because an authenticated tool call obtains current application state at request time; stale embeddings cannot guarantee a transaction's latest status, restoring the ability to let an agent retrieve a live order status from the organization's transactional system.

Incorrect Answers

A:

If the failing function were to personalize repeated interactions without placing all old transcripts in every prompt, this change could be relevant because explicitly persisted, consent-governed preference data supports bounded long-term personalization; transient chat state should not become an uncontrolled permanent profile; the actual symptom is that the assistant summarizes historical order records but cannot report a shipment that changed status minutes ago, which calls for a different repair.

B:

If the failing function were to map queries and reference passages into vectors for semantic retrieval, this change could be relevant because embeddings encode text into vectors used by a similarity search system; they are not themselves a response-generating chatbot; the actual symptom is that the assistant summarizes historical order records but cannot report a shipment that changed status minutes ago, which calls for a different repair.

C:

If the failing function were to answer policy questions using the latest approved internal manuals instead of model recollection, this change could be relevant because a connected retrieval source gives an agent current authorized reference content that can be cited and updated separately from model weights; the actual symptom is that the assistant summarizes historical order records but cannot report a shipment that changed status minutes ago, which calls for a different repair.

Question 36

After implementing a workflow to separate deployment names from model names, the team needs a robust release gate. Which check should it choose?

  1. Exercise the deployment endpoint from a smoke test and check resolved deployment metadata
  2. Run drift detection and compare provisioned configuration with approved definitions
  3. Load-test p95 latency, provisioned utilization, and the configured geography boundary
  4. Check pipeline rejection on a deliberately failing evaluation and verify rollback readiness

Explanation

Correct Answer

A:

The accepted change is to separate deployment names from model names; this evaluation checks whether it can invoke a specific deployed model from application code when a catalog model has multiple deployments, since a model's catalog identity is not necessarily the deployment identifier used by inference requests; deployment references need environment-aware configuration.

Incorrect Answers

B:

This test concerns the ability to reproduce Foundry resources, roles, and deployment settings across test and production environments, for which infrastructure-as-code records resource configuration for reviewable repeatable deployments and reduces portal-only drift; secrets remain in secure stores; it does not establish whether the change to separate deployment names from model names works for the current release.

C:

This test concerns the ability to support high predictable usage while requiring processing within an Azure geography, for which regional provisioned deployments reserve model throughput within an Azure geography where supported; this couples residency with more predictable capacity; it does not establish whether the change to separate deployment names from model names works for the current release.

D:

This test concerns the ability to prevent a newly deployed model or prompt from reaching users before quality and safety checks pass, for which a delivery pipeline can test application behavior, evaluations and deployment health before promotion; versioned artifacts support rollback; it does not establish whether the change to separate deployment names from model names works for the current release.

Question 37

An operator can reproduce this error: A model replacement forces all clients onto a new version simultaneously, with no safe rollback path. Which targeted repair is justified?

  1. Introduce controlled routing between versioned deployments and retain a tested rollback target
  2. Switch the pilot to an eligible Global Standard deployment and monitor metered usage
  3. Enable its managed identity and assign the minimum required resource roles
  4. Validate the model-and-deployment-type availability matrix for the target region before deployment

Explanation

Correct Answer

A:

The fault signal is that a model replacement forces all clients onto a new version simultaneously, with no safe rollback path; the targeted fix works because an application-level rollout control separates client configuration from backend deployment selection, enabling staged exposure and rollback, restoring the ability to move a subset of application traffic to a newly validated model without changing every client.

Incorrect Answers

B:

If the failing function were to serve an unpredictable low-to-medium traffic pilot without geographic processing restrictions, this change could be relevant because global Standard offers pay-per-token inference with flexible global routing; it does not reserve a fixed amount of dedicated throughput; the actual symptom is that a model replacement forces all clients onto a new version simultaneously, with no safe rollback path, which calls for a different repair.

C:

If the failing function were to avoid storing long-lived service keys in source control while authorizing access to Foundry resources, this change could be relevant because managed identity provides an Azure-hosted workload an Entra identity; it still needs the correct data-plane roles and reachable resource endpoint; the actual symptom is that a model replacement forces all clients onto a new version simultaneously, with no safe rollback path, which calls for a different repair.

D:

If the failing function were to avoid selecting a model or deployment type unsupported in the intended Azure region, this change could be relevant because Foundry model availability varies by model, region and deployment type; a catalog listing alone does not mean every combination can be deployed; the actual symptom is that a model replacement forces all clients onto a new version simultaneously, with no safe rollback path, which calls for a different repair.

Question 38

A production AI solution is being scoped for production. The non-negotiable task is to move a subset of application traffic to a newly validated model without changing every client. Which implementation fits?

  1. Use private connectivity and correct DNS resolution
  2. Use managed identity for a production Azure-hosted app
  3. Use deployment aliases or versioned configuration for controlled rollout
  4. Separate deployment names from model names

Explanation

Correct Answer

C:

For the stated requirement to move a subset of application traffic to a newly validated model without changing every client, an application-level rollout control separates client configuration from backend deployment selection, enabling staged exposure and rollback; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to keep Foundry resource access off public paths for a network-restricted application, private endpoints route traffic over approved private networking; proper private DNS and routing are prerequisites for clients to reach the service; this does not directly meet the required outcome to move a subset of application traffic to a newly validated model without changing every client.

B:

For a system whose objective is to avoid storing long-lived service keys in source control while authorizing access to Foundry resources, managed identity provides an Azure-hosted workload an Entra identity; it still needs the correct data-plane roles and reachable resource endpoint; this does not directly meet the required outcome to move a subset of application traffic to a newly validated model without changing every client.

D:

For a system whose objective is to invoke a specific deployed model from application code when a catalog model has multiple deployments, a model's catalog identity is not necessarily the deployment identifier used by inference requests; deployment references need environment-aware configuration; this does not directly meet the required outcome to move a subset of application traffic to a newly validated model without changing every client.

Question 39

After deployment, the team observes: A new prompt version is promoted automatically even though regression tests show more fabricated answers. Which remediation addresses the likely cause?

  1. Declare the missing resource and authorization settings in the reviewed deployment template
  2. Add representative quality and safety evaluations as required checks before production promotion
  3. Inject the correct environment-specific project endpoint and use a permitted identity credential
  4. Introduce controlled routing between versioned deployments and retain a tested rollback target

Explanation

Correct Answer

B:

The fault signal is that a new prompt version is promoted automatically even though regression tests show more fabricated answers; the targeted fix works because a delivery pipeline can test application behavior, evaluations and deployment health before promotion; versioned artifacts support rollback, restoring the ability to prevent a newly deployed model or prompt from reaching users before quality and safety checks pass.

Incorrect Answers

A:

If the failing function were to reproduce Foundry resources, roles, and deployment settings across test and production environments, this change could be relevant because infrastructure-as-code records resource configuration for reviewable repeatable deployments and reduces portal-only drift; secrets remain in secure stores; the actual symptom is that a new prompt version is promoted automatically even though regression tests show more fabricated answers, which calls for a different repair.

C:

If the failing function were to connect a Python application to its intended Foundry project instead of accidentally calling another environment, this change could be relevant because a project-scoped endpoint and appropriate credential select the target project and permitted resources; hardcoded secrets add avoidable risk; the actual symptom is that a new prompt version is promoted automatically even though regression tests show more fabricated answers, which calls for a different repair.

D:

If the failing function were to move a subset of application traffic to a newly validated model without changing every client, this change could be relevant because an application-level rollout control separates client configuration from backend deployment selection, enabling staged exposure and rollback; the actual symptom is that a new prompt version is promoted automatically even though regression tests show more fabricated answers, which calls for a different repair.

Question 40

A development group needs its AI workflow to connect a Python application to its intended Foundry project instead of accidentally calling another environment. What technical approach most directly enables that?

  1. Run evaluated model changes through CI/CD gates
  2. Use private connectivity and correct DNS resolution
  3. Separate deployment names from model names
  4. Use the project's scoped endpoint and identity

Explanation

Correct Answer

D:

For the stated requirement to connect a Python application to its intended Foundry project instead of accidentally calling another environment, a project-scoped endpoint and appropriate credential select the target project and permitted resources; hardcoded secrets add avoidable risk; this is the mechanism that the system must implement.

Incorrect Answers

A:

For a system whose objective is to prevent a newly deployed model or prompt from reaching users before quality and safety checks pass, a delivery pipeline can test application behavior, evaluations and deployment health before promotion; versioned artifacts support rollback; this does not directly meet the required outcome to connect a Python application to its intended Foundry project instead of accidentally calling another environment.

B:

For a system whose objective is to keep Foundry resource access off public paths for a network-restricted application, private endpoints route traffic over approved private networking; proper private DNS and routing are prerequisites for clients to reach the service; this does not directly meet the required outcome to connect a Python application to its intended Foundry project instead of accidentally calling another environment.

C:

For a system whose objective is to invoke a specific deployed model from application code when a catalog model has multiple deployments, a model's catalog identity is not necessarily the deployment identifier used by inference requests; deployment references need environment-aware configuration; this does not directly meet the required outcome to connect a Python application to its intended Foundry project instead of accidentally calling another environment.

Developing AI Apps and Agents on Azure (AI-103) Practice Tests