Task 1.3: Manage, monitor, and secure AI systems

• ARequest a quota increase for the model in that region

  Incorrect

  More quota would hide the problem at extra cost; the deployment isn't short of real tokens.

• BSet the maximum output tokens close to the real summary length

  Correct

  The rate limiter counts an estimate that includes the max_tokens setting when each request arrives, so an oversized value uses up the TPM budget early.

• CTurn on streaming so that tokens are returned as they are generated

  Incorrect

  Delivery style doesn't change how the rate limiter estimates each request.

• DRaise the requests-per-minute limit independently of the TPM value

  Incorrect

  RPM is set in proportion to the TPM assignment; you can't raise it on its own, and requests aren't the bottleneck here.

TPM rate limits use an estimate made when a request is received, built from the prompt and the max_tokens setting. Billing uses the real count afterwards. Right-sizing max_tokens frees the budget.

• AAllow calls to the Foundry resource only from the app's outbound IP addresses

  Incorrect

  Network rules limit where calls come from, but the app would still authenticate with the key.

• BRotate the API key every 30 days by using an Azure Automation runbook

  Incorrect

  Rotation shortens a key's life; it doesn't remove key-based access.

• CUse the app's managed identity with an RBAC role and Entra tokens; disable keys

  Correct

  Keyless authentication uses Microsoft Entra ID: the managed identity gets a role on the Foundry resource, the code requests a token, and local (key) auth can then be disabled.

• DMove the API key into Azure Key Vault and read it each time the app starts

  Incorrect

  Key Vault protects the secret, but the call still uses a shared key, which the rule forbids.

Keyless access means Entra ID tokens plus RBAC. Give the calling service a managed identity, assign it a role on the Foundry resource, acquire tokens for the https://ai.azure.com/.default scope, and remove key auth once no caller needs it.

• AAssign the Reader role to the developer at subscription scope

  Incorrect

  Reader also works, but it grants read access far beyond quota and deployments.

• BAssign Cognitive Services Usages Reader to the developer at subscription scope

  Correct

  Quota comes from the subscription-scoped Usages API, and this role is the least-privilege way to read it.

• CAssign Cognitive Services Usages Reader to the developer on the Foundry resource

  Incorrect

  Resource group, resource and project assignments don't authorize the subscription-level Usages API.

• DAssign Foundry Account Owner to the developer on the Foundry resource

  Incorrect

  That role manages resources and projects; it is broader than needed and isn't subscription-scoped.

Quota visibility is a subscription-level read. Grant Cognitive Services Usages Reader at the subscription; assignments at narrower scopes don't reach the Usages API.

• AWhether the indexer needs a reset so that every document is processed again

  Incorrect

  A reset reprocesses everything, which is expensive and doesn't tell you why particular files failed.

• BThe search service's query latency and throttling metrics for the last hour

  Incorrect

  Query latency describes serving, not whether the PDFs were indexed.

• CThe run's errors and warnings; a run can succeed while some documents fail

  Correct

  A run is marked Success when failed items stay under the Max failed items setting, and document-level problems are listed in its errors and warnings.

• DThe semantic ranker configuration on the index that the app queries

  Incorrect

  Ranking only orders documents that are in the index; missing documents point to ingestion.

Indexer health has two layers: the run status and the per-document results. Open the execution details or call Get Indexer Status to read the errors and warnings before changing anything.

• AThe created timestamp on each of the responses

  Incorrect

  A timestamp doesn't show remaining capacity.

• BThe finish_reason value on each choice it gets back

  Incorrect

  It says why generation stopped for that call, not how much of the per-minute limit is left.

• CThe x-ratelimit-remaining-* response headers

  Correct

  Rate limit headers report remaining tokens and requests so clients can monitor usage and avoid 429s proactively.

• DThe model version field returned in each response

  Incorrect

  The version says which model answered, not how much quota is left.

Use the x-ratelimit-limit, -remaining and -reset headers to pace requests before the limits are reached.

• AAdd several Key Vault connections, one for each of the projects

  Incorrect

  Key Vault connections are limited to one per Foundry resource.

• BDelete the Key Vault later to rotate all of the secrets at once

  Incorrect

  Deleting the underlying Key Vault breaks the Foundry resource.

• CAdd it at any time; the existing secrets migrate automatically

  Incorrect

  Foundry doesn't support secret migration.

• DAdd it while the resource has no other connections yet

  Correct

  The prerequisites require a Foundry resource with no existing connections at the resource or project level.

Bringing your own Key Vault is a setup-time decision: one connection per resource, created before other connections, and the vault must not be deleted while the resource uses it.

• ADiagnostic logging to a Log Analytics workspace

  Correct

  Portal metrics keep 30 days; longer retention and query strings need diagnostic logging, with Log Analytics recommended.

• BA semantic configuration on the main index

  Incorrect

  Semantic settings affect ranking, not logging.

• CAn indexer schedule that runs more often

  Incorrect

  Indexer schedules affect ingestion, not query logs.

• DMore replicas so metrics are kept longer

  Incorrect

  Replicas don't change how long logs and metrics are kept.

Search monitoring has two layers: built-in metrics in the portal and resource logs through diagnostic settings for longer retention and detailed analysis with Kusto.

• AA Key Vault connection for all of the project's secrets

  Incorrect

  Key Vault protects secrets; it doesn't isolate network traffic.

• BPublic network access enabled from all networks

  Incorrect

  That widens access instead of restricting it.

• CA managed virtual network for the Foundry resource

  Correct

  A managed network gives a secure default that restricts what agents can access and helps prevent data exfiltration.

• DAn API key rotation schedule for the resource

  Incorrect

  Rotation affects credentials, not network paths.

Network isolation choices range from bringing your own virtual network to a managed virtual network that Foundry operates; check the supported regions and limitations.

• AA hard spending cap on the Foundry resource that stops calls

  Incorrect

  Azure OpenAI doesn't currently offer hard spending limits; budget notifications can only start automation that you build yourself.

• BA Cost Management budget on that resource group with an alert

  Correct

  Budgets track spending against a limit at the scope you choose and send alerts when costs approach or exceed the thresholds you set.

• CCost analysis filtered by the project tag on the resource

  Incorrect

  Cost analysis shows spend and trends, but viewing it doesn't notify anyone when a threshold is reached.

• DA daily export of the cost data to a storage account

  Incorrect

  Exports support later analysis in tools such as Excel or Power BI; they don't alert on a threshold.

Estimate with the pricing calculator, track with cost analysis and set budgets with alerts at the subscription or resource group scope. A budget warns; it doesn't stop usage on its own.