Microsoft AI-103 Practice Test 18
Question 170
A production assistant must remain source-faithful, avoid unsafe tool behavior, and prove a prompt update generalizes. Which THREE independent tests are appropriate? Choose THREE.
- Score groundedness against source references, not only fluency
- Judge tool-call safety only by whether the final text sounds polite, without inspecting actual attempted API actions.
- Assess output length and usability against a mobile-oriented answer rubric
- Use separate representative validation data and regression holdout cases
- Rerun only the prompt-update examples used during development and treat their scores as proof that new user phrasing will generalize.
- Evaluate the agent tool trajectory and policy compliance, not just final text
Explanation
Correct Answers
A:A release gate should score whether generated claims are supported by the referenced sources, but that is only one dimension. The scenario also requires separate held-out regression evidence and tool-trajectory compliance checks to assess safety and overall reliability.
D:Evaluating only on examples used to tune a prompt creates overfitting and misleading confidence.
F:A successful final answer does not justify unauthorized intermediate actions.
Incorrect Answers
B:Output politeness does not establish that an agent called only authorized tools with valid arguments. Tool trajectory and policy checks are required to assess unsafe actions.
C:Response usefulness includes meeting interface constraints as well as factual correctness. Here, however, the stated objective is to evaluate grounded answers, safe actions and out-of-sample quality; this option focuses on test answer concision for a specified interface.
E:The development set is not an independent holdout. Representative unseen regression cases are necessary to detect prompt overfitting across new inputs.
Question 171
A company must enforce uniform tagging, data residency and allowed regions for Foundry resources. What governance feature fits?
- Apply scoped Azure Policy definitions and initiative assignments
- Plan resilient grounding-data storage and test restoration for the required failure scope
- Give each calling workload its own managed identity and scoped role
- Provision additional stateless controllers and assume their availability pattern enforces region, tag and residency rules.
Explanation
Correct Answer
A:Azure Policy can audit or constrain resource properties across the governed Azure scopes. This directly supports the need to enforce declarative resource governance requirements, unlike a mechanism primarily intended to maintain usable grounding data after a storage failure.
Incorrect Answers
B:Redundant controllers cannot answer grounded questions if their entire document store disappears; durability, recovery and data-location requirements must be designed for the data dependency too. That capability is appropriate for maintain usable grounding data after a storage failure, not this separate task of finding how to enforce declarative resource governance requirements.
C:Distinct workload identities permit least-privilege access without distributing a shared secret. Here the needed outcome is to enforce declarative resource governance requirements; an option designed to isolate each calling service's access rights would not directly provide it.
D:High availability is not governance compliance. Apply the intended Azure Policy assignments at appropriate scopes and validate that disallowed regions or resource configurations are denied or audited.
Question 172
An order agent retries a payment tool after a timeout and accidentally charges twice. Which distributed-systems safeguard is required?
- Use an idempotency key and verify prior transaction state before repeating the side effect.
- Repeat the payment request immediately with a fresh random transaction identifier.
- Trust the client's timeout to mean the provider never processed the first payment.
- Retry without querying prior payment state whenever a response is slow.
Explanation
Correct Answer
A:Timeouts do not prove the first call failed; idempotent execution prevents duplicate business transactions.
Incorrect Answers
B:A new identifier defeats deduplication and may result in a second charge.
C:Network timeout leaves transaction outcome uncertain; it is not evidence of failure.
D:Blind replay of a side effect can duplicate a charge after a delayed first response.
Question 173
An AI app must write logs to an independent workspace that operations can access after a failed deployment. How should diagnostics be designed?
- Write all diagnostic events only to the application's ephemeral local container filesystem and destroy it during redeployment.
- Load test end-to-end model calls, tool latency and downstream concurrency limits
- Give each calling workload its own managed identity and scoped role
- Route service diagnostics and traces to managed centralized monitoring resources
Explanation
Correct Answer
D:External observability makes post-failure investigation possible even when application instances are replaced. The decisive difference is whether the design can retain operational telemetry independent of app instances; the competing function serves to isolate development and production resources.
Incorrect Answers
A:Local ephemeral logs can vanish with the failed instance or deployment. Operational diagnosis requires governed, independently retained service diagnostics and correlated traces in a suitable monitoring destination.
B:AI scalability depends on model throughput and tool dependencies, not only the web server instance count. Here the needed outcome is to retain operational telemetry independent of app instances; an option designed to validate full pipeline capacity under a peak would not directly provide it.
C:Distinct workload identities permit least-privilege access without distributing a shared secret. Its primary purpose is to isolate each calling service's access rights; the stated constraint instead demands a way to retain operational telemetry independent of app instances.
Question 174
A Python app currently contains handcrafted raw HTTP for every inference call and a team wants supported client abstractions. What should be introduced?
- Keep handcrafted HTTP requests and add a permissive optional-field JSON parser instead of using the supported SDK client.
- Use the supported Foundry SDK client for the deployed model and its contracts
- Reuse appropriately managed clients and connections across requests
- Consume a supported streaming inference response and assemble events in order
Explanation
Correct Answer
B:SDK clients centralize request and response handling for supported inference workflows. Here the relevant capability must let the team consume supported model APIs from Python. An alternative targeting remain compatible with optional connector fields solves another problem.
Incorrect Answers
A:A tolerant parser may stop failures on additive metadata, but it leaves the app managing every HTTP call and authentication detail manually. The stated need is a supported client abstraction, not just flexible JSON parsing.
C:Connection/client reuse reduces repeated initialization overhead without weakening authentication. That capability is appropriate for avoid repeated per-token client setup, not this separate task of finding how to consume supported model APIs from Python.
D:Streaming emits incremental content that a client can progressively display while handling completion and errors. Here the needed outcome is to consume supported model APIs from Python; an option designed to render model generation incrementally would not directly provide it.
Question 175
An AI application previously reached a Foundry API publicly; security now requires traffic over an approved virtual network address. What infrastructure is required?
- Configure the supported private endpoint, VNet DNS resolution, and service-side public-access restrictions.
- Allow inbound internet traffic by opening a public firewall rule to the model.
- Create only a private DNS zone without establishing a private service endpoint.
- Assign the calling identity Owner without changing networking.
Explanation
Correct Answer
A:Private Link connectivity depends on approved endpoint association and private name resolution; an allow rule alone does not automatically make model traffic private.
Incorrect Answers
B:A public allow rule fails the requirement to use a private VNet path.
C:DNS alone does not create the network interface and approved Private Link connection.
D:RBAC permissions do not make a public model endpoint privately reachable.
Question 176
A retrieval application has two data connectors that emit differently named source fields. What application boundary helps?
- Consume a supported streaming inference response and assemble events in order
- Provide an explicit supported noninteractive workload-identity credential chain in CI
- Discard records whenever two data connectors use different source-property names instead of mapping them to a shared schema.
- Normalize connector outputs to a common typed evidence schema
Explanation
Correct Answer
D:A typed normalization layer lets downstream prompts and citations treat distinct connectors consistently. This directly supports the need to harmonize evidence from multiple connectors, unlike a mechanism primarily intended to render model generation incrementally.
Incorrect Answers
A:Streaming emits incremental content that a client can progressively display while handling completion and errors. The function serves to render model generation incrementally, whereas the actual requirement is to harmonize evidence from multiple connectors.
B:Developer sign-in state is not automatically available to a pipeline; CI needs a supported identity. That capability is appropriate for authenticate Foundry SDK within CI, not this separate task of finding how to harmonize evidence from multiple connectors.
C:Dropping every incompatible record sacrifices source coverage. Normalize connector-specific fields into an explicit application model so retrieval and citation logic can consume both source shapes.
Question 177
A healthcare team is designing a Foundry deployment that cannot expose public ingress, must maintain zone resilience, and must constrain regions by policy. Which THREE architectural controls directly address these requirements? Choose THREE.
- Load test end-to-end model calls, tool latency and downstream concurrency limits
- Apply scoped Azure Policy definitions and initiative assignments
- Use separate appropriately governed projects and resource scopes for environments
- Give each calling workload its own managed identity and scoped role
- Disable public access to the model endpoint and route clients over supported private endpoints with private DNS.
- Distribute critical stateless controllers across fault-isolated instances
Explanation
Correct Answers
B:Azure Policy can audit or constrain resource properties across the governed Azure scopes. For the stated architecture, Azure Policy constrains permitted regions, supported private endpoints with disabled public network access close the public-ingress path, and fault-isolated controller instances address availability. These safeguards target different requirements; deployment performance testing or extra service identities alone cannot replace any of them.
E:Disabling public ingress and using private connectivity directly addresses the confidential model-access requirement.
F:Zone-separated stateless controllers help the service tolerate one zone loss, but a secure healthcare deployment also needs private network ingress and enforceable governance. Test the controller and data dependencies under zone failure without relaxing policy or exposing a public endpoint.
Incorrect Answers
A:AI scalability depends on model throughput and tool dependencies, not only the web server instance count. That is useful for validate full pipeline capacity under a peak, but it does not directly accomplish the requested need to implement private networking, zone resilience and region policy.
C:Environment separation limits accidental cross-environment changes and supports distinct access and cost controls. That mechanism concerns isolate development and production resources, whereas this case tests how to implement private networking, zone resilience and region policy.
D:Distinct workload identities permit least-privilege access without distributing a shared secret. It would help when attempting to isolate each calling service's access rights; the current failure mode is different: implement private networking, zone resilience and region policy.
Question 178
A connector sometimes responds with a new optional field and the existing parser crashes. What should a robust client do?
- Reject every connector response that contains a new optional JSON property, even when all required fields are valid.
- Provide an explicit supported noninteractive workload-identity credential chain in CI
- Validate tool-result types and treat returned content as untrusted evidence
- Parse documented required fields and tolerate compatible optional additions
Explanation
Correct Answer
D:Defensive schema handling prevents harmless additive metadata from breaking integrations. The decisive difference is whether the design can remain compatible with optional connector fields; the competing function serves to harmonize evidence from multiple connectors.
Incorrect Answers
A:A forward-compatible client should validate required fields while accepting documented or compatible optional additions. Rejecting harmless new properties turns a nonbreaking API response into an avoidable outage.
B:Developer sign-in state is not automatically available to a pipeline; CI needs a supported identity. Here the needed outcome is to remain compatible with optional connector fields; an option designed to authenticate Foundry SDK within CI would not directly provide it.
C:Even a trusted API can return malformed or attacker-controlled content; validation precedes prompting. Its primary purpose is to enforce typed untrusted tool-result handling; the stated constraint instead demands a way to remain compatible with optional connector fields.
Question 179
An AI solution needs a stable model version in production while developers test a replacement. What deployment practice helps?
- Define least-privilege per-department service identities and access scopes
- Version model identifiers, deployment settings, tool schemas and prompt artifacts together
- Move the production endpoint to any region with spare quota without pinning the serving model version.
- Expose separate versioned deployments and promote through controlled environments
Explanation
Correct Answer
D:Keeping production and candidate deployments separate supports comparison and rollback. This resolves the specific need to stage model replacements independently of production, while a measure that helps separate departmental serving permissions addresses a different requirement.
Incorrect Answers
A:Shared serving infrastructure does not imply shared authorization to every model, file or tool. Its primary purpose is to separate departmental serving permissions; the stated constraint instead demands a way to stage model replacements independently of production.
B:Generation behavior depends on more than code; recording AI configuration helps reproducible rollbacks. That is relevant to reproduce the deployed AI configuration, but the present scenario requires an ability to stage model replacements independently of production.
C:Available capacity does not guarantee that a region offers the chosen model/version or preserves production behavior. Version deployments independently and test their cutover.
Question 180
A customer asks a model to update an order, but the user is authenticated only for read-only support. What should block the update?
- Have the model reinterpret the user's read-only role as write permission.
- Execute the update first and check the customer's permissions afterward.
- Enforce the caller's server-side tool permission scope before accepting the mutation.
- Hide the update confirmation in the final reply and skip backend role checks.
Explanation
Correct Answer
C:A natural-language request cannot expand the caller's role; the API must reject unsupported side effects even when the model asks.
Incorrect Answers
A:Model text cannot confer the authorization required by the order-update backend.
B:Checking permission after executing an order update allows an unauthorized mutation to occur. The tool must validate the authenticated caller's write scope against the target order before performing any side effect.
D:Hiding the response to an unauthorized order update conceals activity rather than enforcing access control. The backend needs caller-specific authorization checks before modifying the order, independently of the model's message.