AI-103 — frequently asked questions

While Managed Identity is more secure, it requires additional configuration steps like assigning roles and enabling the identity, which contradicts the requirement to minimize administrative effort.

It is a unique credential provided by the service that allows applications to authenticate and access API endpoints directly without needing complex OAuth flows.

A single-service resource has its own key/endpoint for one API (e.g., only Vision). A multi-service resource ('CognitiveServices') shares one key/endpoint for multiple APIs, simplifying management and billing.

PUT is used to create a new resource or replace an existing one at a specific URI. PATCH is used for updates, and DELETE for removal.

Use Classification when you want to label the entire image content (e.g., is this a cat?). Use Object Detection when you need to locate and count multiple items within an image using bounding boxes.

Binary is for two mutually exclusive outcomes (Yes/No). Multiclass is for three or more distinct categories (Cat/Dog/Bird), even if currently only two exist.

The toolset parameter only makes the tool available to the agent. Without tool_choice, the LLM can still decide not to call it if it believes it knows the answer.

response_format controls how the final answer is structured (e.g., JSON schema), while tool_choice controls which external functions the agent is allowed or forced to execute.

NER returns named entities such as locations, people, and organizations; common nouns like tour and visit are not extracted by this operation.

Only if the function performed extractive summarization or echoed the input text; entity recognition returns entity spans, not the whole string.

Model Catalog is used for discovering and evaluating public/pre-trained models via leaderboards. Model Registry is for tracking versions of models you have trained or customized.

Serverless deployments abstract the infrastructure, meaning the vendor manages the hardware and storage of model weights, whereas container-based deployments give you control over the storage location.

GPT-4 generates text responses. It does not output the raw numeric vectors required for direct similarity calculations without significant custom post-processing.

They convert text into high-dimensional vectors that capture semantic meaning, enabling efficient similarity searches and clustering.

'LLM' commonly stands for both 'Large Language Model' in tech and 'Master of Laws' in legal education. Without context, the LLM cannot determine which definition the user wants.

Context steering adds explicit constraints to the prompt, guiding the model's attention mechanism to focus on relevant tokens and domain-specific knowledge, thereby reducing irrelevant outputs.

The model name defines the architecture, but the deployment name specifies the exact versioned instance running on the server.

No, the endpoint URL is required to locate the resource, and the deployment name is needed to target the specific model instance.

Azure Machine Learning offers more flexibility but requires significantly more development effort (coding, pipelines) compared to the low-code approach of Custom Vision.

Yes, Custom Vision models can be deployed to endpoints that process images or video frames in real-time to detect objects or classify images.

Because 'required' only forces the use of any tool. If multiple tools exist, the agent could pick the wrong one. The question requires the specific MCP tool.

It means the outcome is predictable and fixed. Using the tool type ensures the exact same tool is called every time, preventing the base model from skipping it.

Bearer tokens are for dynamic tokens (OAuth/JWT). API keys are static strings usually placed in a custom header, best handled by the apiKey type.

No. Defining the parameter only describes the request. You must link it to a security scheme so Foundry knows to fetch the value from your connection.

A model cascade chains multiple language models sequentially, using a small/cheap model for simple queries and escalating complex ones to larger models.

Smaller models may fail to provide accurate responses for complex reasoning tasks, leading to degraded quality despite lower costs.

HTTP 429 means 'Too Many Requests', indicating you have hit a rate limit or resource cap (Provisioned Utilization), not that the model is unavailable.

Check the Request/Response logs to see if the retrieval step failed to find documents or if the model simply chose not to generate an answer.

Image Moderation scans pixels for inappropriate visuals (violence, nudity). It cannot read or analyze text hidden inside an image for semantic malicious intent.

Yes, when text is extracted from images (via OCR or multimodal processing), Prompt Shields treats that text as document input and scans it for injection patterns.

Use content safety filters, toxic content detection, and input validation pipelines to detect and block unsafe images or prompt injections before they reach the model.

It is a feature that scans model outputs for copyrighted text or code to prevent intellectual property violations, not for input security.

A RAG evaluator assesses the entire pipeline including retrieval quality. Groundedness specifically checks if the text output is faithful to the context, which is the exact requirement here.

No. Guardrails typically filter input/output for safety or policy violations before generation or block unsafe queries, rather than evaluating the factual correctness of the generated response.

Few-shot examples guide style and reasoning patterns but cannot reliably enforce negative constraints or prevent off-topic hallucinations.

No, they control randomness and creativity. They do not restrict the factual domain or ensure the model stays on topic.

Max_tokens limits length, not selection. If the model chooses to omit content, more tokens won't force it back.

Improve prompts, add explicit instructions to include all clauses, or use structured output schemas to enforce completeness.

Video generation is computationally intensive and asynchronous. The API returns a job ID immediately, so you must poll the status endpoint to know when the video is ready.

Attempting to use or download the video before the status is 'completed' will result in errors or missing assets, as the file does not yet exist on the server.

The scenario explicitly states that API keys must not be used for security and compliance reasons. You must use Entra ID-based authentication instead.

It provides a default way to authenticate to Azure services by trying multiple credentials in sequence, integrating with Microsoft Entra ID for secure access.

Use '--api-properties' and pass a JSON object containing '"keySource": "Microsoft.KeyVault"'.

You must specify 'OpenAI' as the service type to enable generative text capabilities.

Precision is calculated as True Positives divided by the sum of True Positives and False Positives. Check the graph for these specific counts.

It means there are no false positives; every object the model identified was actually present in the image.

Computer Vision analyzes visual content (objects, faces), while Document Intelligence extracts structured text and layout from documents like PDFs using advanced OCR.

Yes, but Azure AI Language provides a dedicated, cost-effective, and standardized sentiment analysis API specifically designed for this task without needing prompt engineering.

Read API extracts linear text but loses layout structure. Article layout interpretation would require significant custom coding, violating the 'minimize development effort' constraint.

Document Intelligence extracts and structures text/content from documents. Image Analysis identifies visual objects, tags, and descriptions, not textual content.

An NSG filters packets at subnets or NICs in vnet1 but does not define which virtual networks the Azure Language resource accepts; that allow-list is set on ta1's networking settings.

No. A Language container runs the service locally in a container and does not configure the cloud ta1 resource's inbound virtual network rules.

It generates safety instructions for a model's system prompt; it does not submit sample content through the CS1 content filter or report threshold-level results.

Use the Studio's moderation/test experience or call the Content Safety API with sample inputs and outputs to evaluate severity thresholds and blocklist behavior.

It only reports whether text may reproduce protected or copyrighted material. It never exercises the agent's hate, violence, sexual, or self-harm severity thresholds on your sample questions.

Test sample prompts and responses against the moderation configuration, review severity levels and blocklist hits, then adjust thresholds or blocklists until input and output filtering behave as required.

It reports live moderation traffic, latency, and trends for deployed APIs. It does not let you submit sample questions or compare threshold changes before deployment.

Use the interactive testing or playground experience in Content Safety Studio to submit sample prompts and adjust sensitivity thresholds and blocklists.

That command creates, lists, or deletes firewall and virtual network rules for the account. It changes or reports access settings, not the account's identity, endpoint, or SKU.

No. az account list returns the Azure subscriptions you can access, so it never reveals which Azure AI services resource the pipeline step created.

It requires labeled images, training, evaluation and a custom deployment, which adds effort the question explicitly wants to avoid, and accuracy drops with overlapping people.

It ingests a real-time RTSP video feed from an IP camera, so the user's camera stream must be exposed as RTSP to the Spatial Analysis container or service.

Protected material detection identifies copyrighted material; it is not a prompt-injection or PII-leakage control. Prompt Shields addresses the hidden-instruction attack described in the case.

Prompt Shields targets direct and indirect prompt-injection attacks. For image-embedded text, combine it with multimodal content safety and validation before grounding.

Image Classification assigns a label to the entire image but does not provide spatial coordinates (bounding boxes) indicating where the defect is located.

Choose General (compact) when deploying to edge devices or environments with limited compute power and intermittent connectivity, as these models are smaller and faster.

It adds only a few input tokens and gives GPT-35 Turbo a direct instruction, so you avoid a larger model or a longer engineered prompt.

No. GPT-4-32k costs more and its larger context is unnecessary for explaining a short Fibonacci function; GPT-35 Turbo is sufficient here.

If the images already have the correct tags applied in a previous dataset or via bulk import, you can skip manual labeling. However, for new products, explicit labeling is required to define the new classes.

Custom Vision models are static once published. Adding images only updates the dataset. You must retrain the algorithm to learn the visual features of the new data before publishing the updated model.

A shared key authenticates the resource, not each app, so any holder can call AI1; managed identities let you grant and audit per-app access.

The app places the deployment name in the REST URL path; all three deployments can use the same GPT-3.5 model but have different deployment names.

No. An IP firewall filters traffic but keeps the service's public endpoint active. A Private Endpoint is required to remove public internet access entirely.

Generally no. API keys are typically scoped to the entire search service. RBAC roles allow you to assign permissions to specific indexes or collections.

It needs a trained custom model and does not expose built-in localized tag output, so it adds training and translation work rather than minimizing development effort.

Yes. Its language parameter lets the analyze request return tags in the user’s selected language without a separate translation service.

Azure AI Speech containers are considered gated services. Microsoft requires users to submit an access request form to ensure they meet usage criteria before providing the container image.

Standard speech containers may have different availability, but custom models and specific enterprise features often require explicit approval. Always check the specific documentation for the model type you are deploying.

Generative deployments sample new wording on every call, so the same HR question can produce different phrasing, which breaks the identical-answer requirement.

It can preprocess scanned or complex HR PDFs into clean text, but it has no knowledge base or answer-matching layer, so it cannot guarantee repeatable answers alone.

While LLMs can read text, Azure Vision is purpose-built for high-accuracy OCR in natural scenes. Using Vision minimizes development effort by providing a direct API for extraction without complex prompt engineering.

Azure Language includes Named Entity Recognition (NER) which can identify 'Geographic' entities in text, allowing you to detect well-known locations from the unstructured text output of the Vision service.

Blocking removes the response entirely, leaving the user with an error. Improvement implies the user receives a better, complete summary, which requires regeneration or correction, not just filtering.

Use self-correction loops where the agent reviews its own output against criteria, or adjust the retrieval grounding pipeline to ensure all policy documents are fully ingested and cited.

Input fidelity typically relates to text adherence, not visual anchoring. Without the input image payload, the model cannot preserve visual identity.

It is a safety filter to detect if generated content aligns with the prompt, not a tool for controlling image-to-image transformation fidelity.

IAM controls who can perform actions, not where connections come from. To block external/public network access, you need network-level restrictions.

Action A modifies the Service's allowed networks/VNet settings. Action E modifies general VNet properties. The restriction is applied by configuring the Service (A) to trust the VNet.

It analyzes text for harmful content and allows adjusting sensitivity thresholds for categories like hate speech and violence.

While you can set static thresholds, running tests on sample questions via the Moderate feature helps optimize them based on real-world performance.

Temperature 0 removes sampling randomness, so the model reproduces the same phrasing for the same prompt and the pattern-based validator stops flagging small wording differences.

No — reasoning effort controls how much deliberation the model performs, an independent parameter from temperature, so effort high keeps stability while maximizing reasoning quality.

Each image contains one whole component that simply needs to be sorted into faulty or non-faulty. Object detection is only needed when you must locate and box multiple distinct objects inside an image.

Custom Vision training is supervised: the tags you apply to uploaded images become the labels the model learns from, so an untagged or untrained set cannot produce a model.

User Prompt Attacks originate from direct user input (text/speech). Document Attacks originate from external data sources like uploaded files, images, or third-party APIs that are processed by the application before reaching the model.

Spotlighting marks extracted text from untrusted sources (like images) as 'data' rather than 'instructions'. This prevents the LLM from executing hidden commands embedded within the content.

Object projection stores complex, hierarchical JSON data as individual blob files. Table projection stores extracted data, such as text or key-value pairs, in relational tables for easier querying.

Use File projection only when you need to store normalized, binary image files, such as resized images or thumbnails extracted during the enrichment process.

Tool call intercepts the request sent BY the agent TO the tool, allowing you to filter outgoing parameters. Tool response intercepts the data returned FROM the tool TO the agent, allowing you to filter incoming results.

Yes, you can configure guardrails using Azure AI Content Safety at the tool call intervention point to scan and filter payloads before they are sent to external MCP endpoints.

prebuilt-layout returns raw visual layout analysis — reading order, tables, checkboxes and barcodes — and can render Markdown. prebuilt-documentSearch is the RAG-oriented analyzer that additionally structures content with semantic hierarchy and retrieval-friendly Markdown chunks for search indexes and agent grounding.

No. prebuilt-read is OCR-focused: it extracts text with reading order but drops table structure and semantic hierarchy, producing Markdown that retrieves poorly. Use a RAG analyzer such as prebuilt-documentSearch when the output feeds an index or a downstream agent.

Use it when you need typed field extraction — you define a JSON field schema (names, types, descriptions) and the analyzer returns validated field values. It suits structured data capture like invoices, not free-form document grounding for RAG.

The content to moderate is passed as a plain string with AnalyzeTextOptions(text=comment). The categories parameter is only for a list of custom category names, not the input text.

Use response.categories_analysis.self_harm.severity. It returns an integer from 0 (safe) to 6 (high) for the self-harm category.

Yes. A single call returns hate, sexual, violence, and self_harm severities together, so no separate per-category request is required.

Privacy and security focuses on consent, data minimization, encryption, and access control. A visible notice that AI used the content is transparency; a separate privacy statement covering retention and lawful basis would address privacy.

Provide in-product disclosure that an automated system analyzes uploads for a stated purpose, plus a documented data-use description. For example, notice that uploaded video is analyzed to tag product mentions.

Transparency makes the AI's role visible and understandable to users; accountability establishes human oversight, auditability, and clear ownership of the system's decisions and outcomes.

Summarization reduces content length but returns plain text with no line focus, syllable splitting, read-aloud or spacing controls. It also requires building and maintaining your own reader UI and API integration, violating the minimal-effort requirement.

Launch it from a small JavaScript snippet that embeds the token-secured Immersive Reader iframe, or use the native integration available in SharePoint, Power Apps and Microsoft Teams.

Choose Translator when the barrier is language, serving readers who need content in another language. The dyslexia scenario is about reading mechanics within the user's own language, which Immersive Reader also supports alongside its accessibility features.

Deleting first immediately invalidates the credential the production web app still holds, so every search request fails with 401 Unauthorized until a new key is generated and the app is redeployed. Creating and applying the new key first removes that outage window.

No separate permission change is needed. Query keys in Azure AI Search can only perform read operations such as search and lookup; they cannot modify or delete documents, so staying on a query key keeps the documents collection read-only.

Only the query key involved in the anomalous traffic needs rotation in this scenario, and the same add-then-switch-then-delete order applies. Admin keys are separate credentials with full service rights, so protect them with distinct handling and rotation procedures.

No. The token authenticates the caller to the search service, but authorization at document level still requires a filterable security field plus an OData filter built from the user's group memberships.

It works but consumes index quota, duplicates indexer and datasource configuration, and complicates management at scale. A single index with a filterable group field is the recommended design.

As an OData $filter using a collection operator such as groups/any(g: search.in(g, '<group ids>')), evaluated against a filterable field populated with Microsoft Entra ID group object IDs.

No. It links recognized entities to a knowledge base and assumes a language unless one is supplied in the request options. Detecting language requires a separate detect_language call against the Azure AI Language client.

They come from the underlying data source article, such as Wikipedia, which is also reported in the data_source property. Bing is involved in entity resolution behind the scenes and via bing_entity_search_api_id, but not in the url field.

Each match returned with a linked entity contains a Location object with the character offset and length of the occurrence in the submitted text, so you can locate and replace or link every occurrence.

Virtual network rules use service endpoints so approved subnets reach the resource over the Azure backbone; private endpoints give the resource a private IP inside your VNet. Both restrict access, but virtual network rules are quicker to configure, which matches the minimal-effort requirement.

IPsec is the encryption policy behind site-to-site VPN tunnels and typically implies an on-premises connection. It protects traffic in transit but does not grant or deny access based on the calling Azure process's subnet.

Yes, the same firewall supports IP rules for public address ranges, but those suit on-premises callers. For Azure-hosted processes, virtual network rules are the precise and lowest-effort control.

The container sends authentication and usage/billing telemetry to the Azure AI resource endpoint. The document content itself stays on-premises, but without outbound internet access the container cannot validate the key and will not operate.

Images are published on the Microsoft Container Registry at mcr.microsoft.com (for example, the text analytics and custom NER images). Docker Hub is not the source; only the image is pulled from Azure, never your data.

You must supply ApiKey=<your-key> and Endpoint=<your Azure AI resource endpoint>, typically with --add-port to expose the local REST server. Missing either value prevents the container from starting.

No. The Custom Vision Performance tab reports precision and recall per iteration and per tag at the selected probability threshold; F-score is not presented as a reviewable metric, though you could compute it yourself from precision and recall.

Precision is the share of images the model tagged as a label that actually belong to that label. Recall is the share of all images truly carrying that label that the model successfully identified.

AUC is a threshold-independent metric reported by Azure Machine Learning experiment evaluation, not by the Custom Vision service. Custom Vision evaluates performance per probability threshold using precision and recall.

The Vision Read API is asynchronous: it returns an operation_id immediately and processes the image server-side. get_read_result just fetches the current snapshot, so you must poll it until status is succeeded or failed.

read_result.status is the live state of the OCR job (notStarted, running, succeeded, failed) and is what you poll. read_operation_location is the URL of the operation resource created at submission and never indicates progress.

No. operation_id identifies which asynchronous job to query; removing it would prevent the client from retrieving any result. Poll with the same operation_id inside a delayed loop instead.

With False, the service finishes the analysis and returns both the blocklist matches and the built-in category results. With True, moderation stops as soon as a blocked term is found and returns only that blocklist hit, so category scores are omitted.

After calling client.analyze_text(options), examine the blocklists_match collection on the returned result; each entry identifies the blocklist and the matched text, which you can log, flag, or use to block the message before it reaches the language model.

Yes. ConfidentialTerms must already be created in the Foundry project's Content Safety resource, and the code only references it by name via blocklist_names; the application never uploads or redefines the terms at moderation time.

An embeddings model converts the description into a vector of floating-point numbers for search and similarity; it emits no readable text, so there is nothing to display as a product name. Use embeddings only to rank or retrieve candidates afterwards.

Whisper applies only if the description arrives as audio that must be transcribed first. DALL-E applies if you also need a product image or mock-up; neither can generate the written names themselves.

Yes, as a secondary step: embed existing catalogue names and the generated candidates, then use vector similarity to filter duplicates or rank novelty. The naming itself must still be done by a GPT-4 chat model.

presence_penalty penalizes tokens already seen to encourage variety, which targets repetition and diversity, not coverage of source material. It is also an inference parameter rather than application logic, so it fails the question's implementation constraint.

The app calls the agent to produce a draft summary, then sends a second prompt containing the retrieved policy documents plus the draft asking the model to critique and expand it. The revised summary is returned to the user.

No. Streaming only changes delivery, sending tokens incrementally instead of as one payload. The generated content and its completeness are identical either way.

Microsoft documents the container billing endpoint in its HTTP form, so you pass Billing=http://contoso.cognitiveservices.azure.com even though the portal displays an https:// URI. The container validates this value along with ApiKey against your Azure AI Language resource.

Use the sentiment-specific image: mcr.microsoft.com/azure-cognitive-services/textanalytics/sentiment. Each Language capability (sentiment, language detection, PII, key phrases) has its own image, so a generic image will not serve sentiment requests.

Without Eula=accept the container will not start. Without a valid ApiKey (and a matching Billing endpoint), the container starts but rejects every sentiment request with an authorization error, so verify both before testing.