Part 1 Full Quiz

C is correct. Document Intelligence (formerly Form Recognizer) is designed for extracting structured fields from documents with a consistent layout. A custom template model learns the specific positions of "Patient Name", "Medication", "Dosage" fields in your prescription forms. AI Language NER extracts general medical entities but doesn't understand form structure. AI Vision OCR extracts raw text without field understanding. AI Search enables searching, not structured field extraction.

A is correct. Azure Cognitive Search was renamed to Azure AI Search. QnA Maker was retired and replaced by Custom Question Answering, which is now a feature inside Azure AI Language. These name mappings are frequently tested. Other critical renames: LUIS → CLU (inside Azure AI Language), Form Recognizer → Document Intelligence, Text Analytics → Azure AI Language, Computer Vision → Azure AI Vision.

B is correct. Private endpoints — the mechanism that removes a resource from the public internet — require S0 (Standard) tier and are not available on F0 (Free). Before any private endpoint can be configured, the resource must first be upgraded to S0. Firewall rules (option A) only restrict which IPs can reach the public endpoint; they don't remove the resource from the internet entirely. Managed identity (option C) controls authentication, not network access. Option D is not a valid Azure concept.

B is correct. Azure AI Video Indexer is purpose-built for this use case. It automatically extracts transcripts, keywords, topics, named entities, and face identification from video files at scale — making them searchable without manual transcription. Option A would only provide transcription (audio to text) but not topics or person identification. Option C (Spatial Analysis) analyses live video streams for people movement/counting — not stored video files for content indexing. Option D would only summarise but not provide the rich metadata (keywords, faces) needed for precise search.

B is correct. By using ApiKey=$LANGUAGE_KEY, the actual key value is read from an environment variable at runtime and never appears as plain text in the command or command history. The shell expands the variable before executing the command but doesn't store the key value in history. Option C is wrong — ApiKey is mandatory for all Azure AI containers, even offline ones (they must authenticate to Azure for billing). Option D (a text file) would still expose the key on disk. Option A (private endpoint) controls network access, not command history.

B is correct. A RAG pipeline requires two different model types: (1) an embedding model to convert text to vectors — this is always text-embedding-ada-002 in Azure OpenAI; (2) a chat/completion model to generate responses from retrieved context — either gpt-4o (best quality) or gpt-35-turbo (cheaper). DALL-E generates images, not text responses. Whisper transcribes audio, not text vectors. The embedding model cannot generate conversational responses.

B is correct. Even though the service was renamed from Form Recognizer to Document Intelligence, the pip package name is still azure-ai-formrecognizer. The package name was not updated when the service was rebranded. This is one of the most tested package name traps in the AI-102 exam. Always install azure-ai-formrecognizer when working with Document Intelligence in Python.

B is correct. Azure AI service containers are designed for low-connectivity environments. They store usage telemetry locally and report back to Azure for billing whenever internet connectivity becomes available — they do not require continuous connectivity. However, the Billing endpoint and ApiKey parameters are still required in the docker run command (they're used for authentication when the container does connect). Option A is wrong — containers do bill. Option C is wrong — they operate offline. Option D is not a real Azure offering for this scenario.

C is correct. A User-Assigned Managed Identity is designed exactly for this scenario — it can be assigned to multiple resources simultaneously and managed centrally. No credentials to store (satisfies the security policy). Option A (System-Assigned) creates separate identities per resource, requires separate role assignments for each, and cannot be centrally managed as a single identity. Option B still uses an API key (a credential), even if stored in Key Vault — the policy says no credentials in code OR configuration. Option D involves a client secret (a credential).

B is correct. The Contributor role grants management permissions — creating, modifying, and deleting resources, including changing network settings. However, Contributor does NOT grant access to call the AI service endpoint (option A requires Cognitive Services User role) and does NOT allow assigning RBAC roles to others (option C requires Owner role). This is the classic RBAC trap: management permissions ≠ endpoint access.

B is correct. This is the zero-downtime key rotation procedure. Step 1: Switch all running applications to Key2 (Key2 is valid — no downtime). Step 2: Regenerate Key1 (old Key1 is now invalid, but apps are using Key2 so no impact). Step 3: Switch apps back to the new Key1. Azure provides two keys specifically for this rotation pattern. Option A would cause downtime because apps using the old Key1 break the moment it's regenerated. Option C would delete the service itself — causing downtime. Option D is unnecessary and disruptive.

B is correct. The last item is wrong — Soft Delete and Purge Protection must be ENABLED on the Key Vault when using CMK, not disabled. These protections prevent accidental permanent deletion of the encryption key, which would make all encrypted data permanently inaccessible. Microsoft requires these protections as a prerequisite for CMK. Option A is wrong — standard Azure Key Vault works fine for CMK (Azure Managed HSM is not required). Option C is wrong — the full Key Identifier URI is required.

B is correct. Audit logs (including key regeneration events) must be routed to a Log Analytics workspace via diagnostic settings to be queryable. KQL uses == for equality (not SQL's =) and filters with | where. The AzureDiagnostics table contains audit events. Option A (Storage Account) stores logs for archiving but doesn't support SQL queries. Option C (Application Insights) tracks application-level events in your code, not Azure resource management operations. Option D — metrics are numeric measurements, not event records.

C is correct. Azure Event Hub is the correct destination for streaming logs to external SIEM systems like Splunk. Splunk has an official add-on (connector) for Azure Event Hub that enables real-time log ingestion. The flow is: Azure AI Service → Diagnostic Settings → Event Hub → Splunk. Log Analytics (option A) doesn't support real-time streaming to Splunk. Storage Account (option B) would require polling for new files — not real-time. Application Insights (option D) monitors application code, not service logs, and doesn't have SIEM export functionality.

B is correct. Azure alert severity scale: 0 = Critical (most severe), 1 = Error, 2 = Warning, 3 = Informational, 4 = Verbose (least severe). Lower number = higher severity. This is the opposite of what many people intuitively expect. Severity 0 should be used for conditions that require immediate response — service outages, complete unavailability, or security breaches. It should trigger an action group that pages the on-call team immediately.

B is correct. Azure Cost Management budgets are notification-only. When a threshold is hit, Azure sends email alerts to configured recipients (e.g. at 80%, 90%, 100% of budget). The AI service continues running, API calls continue working, and billing continues to accumulate beyond the budget amount. Azure does NOT automatically throttle, stop, or suspend services when a budget is exceeded. To prevent overspending you must manually disable services or configure automation (e.g. a Logic App that acts on budget alerts).

B is correct. Inclusiveness is violated because the system performs significantly worse for users with non-native accents — it fails to serve a diverse user population equally well. Accountability is violated because the company knowingly shipped a product with this flaw without taking responsibility to fix it or disclose it. While Fairness is related, the primary principle for serving all groups including linguistic minorities is Inclusiveness. While Reliability & Safety is also relevant (lower accuracy = less reliable), the most direct match to "knowingly shipped despite the flaw" is Accountability.

B is correct. This is a classic jailbreak attack — the user is directly attempting to override the AI model's system instructions by asking it to adopt a different persona with no restrictions. This comes from the user's message directly (not from a document), making it a User Prompt Attack. Prompt Shields specifically detect this pattern. Content Filters (option A) screen for harmful content categories (Hate, Sexual, Violence, Self-Harm) — they don't detect instruction-overriding attempts. The content itself may not even be "violent" — the attack is about bypassing restrictions, not the content type.

B is correct. The threshold is 4, and the rule is: block content where score ≥ threshold. Violence score is 2, which is less than 4 (2 ≥ 4 is False), so the content is allowed through. All other categories scored 0, which is also below threshold. The content passes the safety check. "Need more review" or "flag" are not automatic behaviors of the Content Safety SDK — it returns scores and your code decides the action based on your threshold. The code pattern taught was: if result.severity >= SEVERITY_THRESHOLD: block.

A is correct. Breaking down each requirement: (1) Private Endpoint removes the resource from the public internet entirely — IP firewall rules only restrict it, not remove it. (2) Managed Identity authenticates without any stored credentials — API keys and service principals are credentials. (3) Alert Rule + Action Group is the mechanism for sending email notifications on failures — Azure Monitor Metrics alone don't send notifications. (4) Customer-Managed Keys (via Key Vault) lets the company control their own encryption keys. (5) Azure AI Content Safety screens arbitrary user-generated text for harmful content. Note: Content Filters (option C option 5) screen AI model inputs/outputs — but Content Safety is the better fit for pre-processing user inputs before they reach the model.