AI-102 — Frequently Asked Questions

Metric alerts use static thresholds and cannot detect context-dependent anomalies like sudden spikes relative to historical trends, which is required for 'atypical' detection.

No, it is a managed API that handles the statistical modeling automatically, allowing developers to integrate detection via simple REST calls or SDKs.

Azure AI Video Indexer does not support indexing videos from third-party streaming services like YouTube due to policy restrictions and technical incompatibilities.

While older documentation cited a 2 GB limit, modern implementations support larger files via direct URLs. However, extremely large files may still face timeouts, making efficient linking crucial.

IAM controls who has permission to use the API keys, but it does not filter the source IP address or network. External attackers with valid keys could still access the service.

Service Endpoints optimize routing and add context to traffic. For this specific question, the 'modify virtual network settings' on the resource side refers to adding subnet IDs to the allowed networks list, which works with Service Endpoints.

Object detection finds specific items inside an image, not the overall image type. Image Type Detection classifies the whole image as a photo, drawing, etc.

Yes, the Image Descriptions feature in Azure AI Computer Vision uses AI to generate human-readable captions describing the visual content.

Text Analytics focuses on sentiment, key phrases, and language detection. It does not support the custom intent classification model training or querying found in Conversational Language Understanding.

The azure-ai- packages are the newer, unified SDKs for Azure AI services. The older azure-cognitiveservices- packages are legacy and may not include the latest features for services like CLU.

Pattern.any extracts variable-length entities inside patterns, but it does not map synonyms like JFK to John F. Kennedy International Airport; list entities do that without extra utterances.

Yes, because each list item can have many synonyms, so one labeled example can cover numerous variations without adding separate utterances.

Priming sets the overall context, persona, or tone (e.g., 'act as a teacher'). Chain of thought explicitly instructs the model to break down its reasoning step-by-step to solve complex problems.

No. Asking for an explanation with a specific audience constraint (like 'explain to a 4-year-old') is priming. Chain of thought requires explicit instructions to show the logical steps taken.

While RAG improves factual grounding, the question asks about model parameters. Temperature directly controls the randomness of the generation process itself.

Top P represents cumulative probability. Setting it to 0 is invalid as it excludes all tokens; it must be a value between 0 and 1.

Custom Vision requires you to label training images and train a model, which adds development effort. The question demands minimal development effort, and Custom Vision has no prebuilt adult-content classifier.

Yes. Content Safety analyzes images and returns harm category severity scores you can use to block uploads, making it a complete solution — as is Azure AI Vision's adult/racy detection.

It only detects copyrighted or protected text, so it returns no severity scores for hate, violence, sexual or self-harm categories that Azure OpenAI content filters evaluate.

Use the moderation / Analyze Text test flow on sample questions to read harm category severity levels, then adjust the content filter configuration applied to the Azure OpenAI resource.

No, the Bot Framework Emulator works independently of Azure Bot Service registration, making it ideal for local development.

While technically part of the curriculum, many experts note it is being phased out in favor of Azure OpenAI and GPT-based solutions.

The SDK routes traffic based on the Deployment Name, which points to a specific instance/version. The underlying model name is metadata, not the routing key.

The endpoint contains the Azure region and resource name (e.g., https://<resource>.openai.azure.com). It identifies the specific Azure OpenAI resource instance.

Virtual network rules directly filter which subnets may reach the Language service, while a gateway only provides connectivity and adds network infrastructure you must manage.

No. IPsec protects traffic inside an existing tunnel; it does not authorize or block callers of the service, so it cannot meet the access requirement.

Time Series Insights is for visualization and exploration. For automated anomaly detection in AI scenarios, Anomaly Detector is the dedicated service.

Yes, Azure AI Anomaly Detector supports multivariate anomaly detection, allowing it to analyze relationships between multiple sensors simultaneously.

Face Service detects facial landmarks and attributes, not semantic intent like 'asking a question'. It cannot distinguish between talking and asking.

No. While the input is video, the task is audio-based. Using Speech-to-text minimizes development effort by leveraging existing transcription APIs.

az account list returns subscription information, not specific resource instance details like an AI service account.

Yes, while Microsoft rebranded to Azure AI Services, the CLI group remains cognitiveservices in many versions.

No. The custom language model accepts plain text (.txt) files containing the terms, either as a list or in full sentences.

Video Indexer's upload step only supports text files; PDF, XML, and XLS are not valid formats for customizing the language model.

operation_id identifies the async read operation, so get_read_result needs it to fetch the correct OCR results; removing it does not make the call wait for completion.

read_operation_location is the polling URL, not a status. read_results.status tells the code whether OCR is still running or has succeeded.

Yes. MOV is a supported container, the 980 MB file is far below the 2 GB device-upload limit, and the clip's length stays within the service's duration rules.

Both exist: 2 GB is the maximum per file when you upload from a device or the website, while ingestion from a URL supports files up to 30 GB.

A Language resource exposes only the Language service, so it cannot also host the Decision API under the same endpoint and credential. The multi-service resource is required to combine both APIs.

The portal now brands it Azure AI services, but the multi-service resource still provides a single key and endpoint for multiple Azure AI APIs.

Entity linking only resolves mentions against a public knowledge base, so internal codenames and many private product names return no match. Custom NER instead learns those exact terms from your labeled examples.

No. Azure AI Language handles model training, hosting, and the prediction endpoint; your main effort is labeling example sentences and retraining when products or codenames change.

It is more secure in production but requires enabling the identity, assigning an RBAC role and managing certificate rotation, which is extra administrative effort the question rules out.

An OAuth token is the output of Entra ID or managed identity authentication, so you would still configure that identity first; the Cognitive Services resource natively issues subscription keys instead.

Because a multi-service key can be used with several services and regions, Translator needs Ocp-Apim-Subscription-Region to route the request to Translator1's deployment.

Yes. The Translator v3 translate operation expects Content-Type: application/json when you send a JSON request body, in addition to the subscription key and region headers.

Yes, the Moderate endpoint is available via REST API and SDKs, but Content Safety Studio provides a UI for easier testing and configuration tuning.

CS1 refers to the Azure AI Content Safety resource instance used to check input and output for objectionable content.