Troubleshooting Scenario Questions

Change the system prompt to ignore safety.

Why wrong: System prompts cannot override platform-level safety policies.

Increase the model temperature.

Why wrong: Temperature does not affect content filtering.

Disable Content Safety for the deployment.

Why wrong: Disabling safety is against best practices.

Increase the Azure OpenAI model temperature parameter to 1.5

Why wrong: Temperature controls generative output randomness, not document retrieval quality.

Configure Azure Key Vault access policies for the search service

Why wrong: Key Vault access manages cryptographic keys, not search relevance.

Restart the client web browser session

Why wrong: Browser sessions do not affect server-side agent tool execution logic.

Increase the Azure DNS TTL value for domain resolution

Why wrong: DNS TTL settings do not fix incorrect tool argument formatting or API failures.

Double the model temperature to 2.0 to accelerate reasoning speed.

Why wrong: Temperature does not impact execution speed or timeouts.

Delete all vector indices to force local memory processing.

Why wrong: Deleting vector indices breaks RAG functionality entirely.

Deleting and recreating the entire resource group.

Why wrong: This is a destructive and unnecessary step.

Increasing the model temperature to see if the error persists.

Why wrong: Changing model parameters is not a debugging technique for flow logic.

Increase the vector index size

Why wrong: Index size does not affect reasoning quality.

Use a smaller base model

Why wrong: Smaller models typically have worse reasoning capability.

Disable streaming response

Why wrong: Streaming is purely for UX.

The system prompt length.

Why wrong: System prompt length has negligible impact on search retrieval latency.

The content safety threshold.

Why wrong: Safety filtering is performed post-retrieval.

The model temperature.

Why wrong: Temperature is a model hyperparameter, not a search parameter.

Use the Read API instead

Why wrong: Read API does not provide structural tables.

Disable the layout model

Why wrong: This is not possible.

Change the resource region

Why wrong: Region does not affect extraction logic.

The model weights have become corrupted and require redeployment.

Why wrong: Model corruption would typically result in HTTP 500 errors, not rate-limiting HTTP 429 errors.

The Azure subscription has run out of monthly financial budget.

Why wrong: Exceeding budget alerts do not automatically trigger HTTP 429 throttling errors from the service.

The Private Endpoint DNS resolution has failed.

Why wrong: DNS resolution failure results in connection timeout or name resolution errors (e.g., DNS_PROBE_FINISHED_NXDOMAIN), not HTTP 429.

Listen for incoming webhook notifications on an Azure Event Grid subscription attached to the text analytics resource.

Why wrong: While Event Grid can be used for some services, standard LRO batch job retrieval requires polling the operation-location URI unless specifically configured otherwise.

Query the Azure Monitor metrics API for job completion status.

Why wrong: Job results and execution status are retrieved via the operation-location API, not Azure Monitor metrics.

Repeatedly send synchronous POST requests to the same document endpoint until a 200 OK is returned.

Why wrong: Synchronous endpoints timeout on large batch payloads; asynchronous jobs require polling the operation-location URL.

Lower the confidence score threshold to 0.01.

Why wrong: Changing inference thresholds does not fix model underfitting.

Reduce the training dataset size by 90%.

Why wrong: Reducing training data will worsen underfitting.

Switch the resource pricing tier from Standard to Free.

Why wrong: Pricing tiers do not change model training algorithms.

Run the Azure Advisor cost recommendation analyzer to optimize resource pricing tiers.

Why wrong: Cost recommendations do not troubleshoot network connectivity issues.

Use the Azure Resource Manager template visualizer to inspect regional resource dependencies.

Why wrong: Template visualizers display deployment hierarchies, not real-time network path diagnostics.

Increase the timeout threshold in the HTTP client configuration.

Why wrong: Increasing the timeout hides the latency issue rather than fixing the cross-region network delay.

Enable Azure ExpressRoute between West US and Southeast Asia.

Why wrong: ExpressRoute connects on-premises datacenters to Azure, not between Azure regions.

Switch from REST API to gRPC protocol.

Why wrong: Azure AI Vision REST API does not support gRPC.

In the styles collection under the isHandwritten property.

Why wrong: The styles collection describes visual styling characteristics, not extracted field values.

In the paragraphs collection under the boundingRegions array.

Why wrong: Paragraphs contain raw text blocks, not typed, normalized field values.

In the analyzeResults object under the content property.

Why wrong: The content property contains the full concatenated string of all extracted text.

Increase the probability threshold in the prediction API call.

Why wrong: While threshold adjustment filters results, fixing the root cause with better negative training data is the correct machine learning remediation.

Remove all tagged bounding boxes and retrain as a multi-class classification model.

Why wrong: Converting object detection to classification will lose spatial location data.

Apply aggressive data augmentation to duplicate existing positive images.

Why wrong: Duplicating positive images does not fix false positives caused by lack of negative training context.