Microsoft Azure AI Engineer Associate AI-102 (AI-102) — Questions 226–300

Why this answer

How to eliminate wrong answers

Option A is wrong because extractive summarization does not limit output to the first sentence based on document length; it selects sentences based on relevance scores, and the number of sentences is controlled by the `sentenceCount` parameter. Option B is wrong because abstractive summarization is a different capability of the Azure AI Language API (using a different endpoint or model), while this request specifically targets extractive summarization, which copies sentences verbatim. Option C is wrong because although sentences are ranked by confidence score internally, the API returns them in the original document order by default, not sorted by score; you would need to explicitly set `sortBy` to `'Rank'` to change this behavior.

Why this answer

Why this answer

How to eliminate wrong answers

Option B is wrong because Azure AI Computer Vision image enhancement does not improve OCR accuracy for Document Intelligence; the service already applies its own preprocessing, and external enhancement may introduce artifacts. Option C is wrong because retraining the model with more table examples will not fix the root cause if the OCR step cannot correctly extract text from low-quality images; the model relies on accurate OCR input. Option D is wrong because using a higher resolution scanner is a hardware solution that may not be feasible for existing PDFs and does not address the immediate diagnostic need; the first step should be software-based verification of OCR output.

Why this answer

How to eliminate wrong answers

Option A is wrong because prebuilt NER models only recognize general entity types (e.g., person, location, organization) and cannot be remapped to extract highly specialized scientific entities like gene or protein names without additional training. Option C is wrong because Azure Logic Apps calling the Text Analytics API would still rely on prebuilt NER capabilities, which cannot extract the specialized entities required, and post-processing would be inefficient and error-prone. Option D is wrong because training a custom NER model for genes while using prebuilt NER for chemicals is inconsistent—prebuilt NER does not recognize chemical compounds in a specialized scientific context, and this hybrid approach would require separate handling and likely reduce accuracy.

Why this answer

Why this answer

How to eliminate wrong answers

Option A is wrong because Pay-as-you-go deployments rely on Azure's global rate limits and do not support automatic scaling; they can throttle requests during traffic spikes, leading to failures. Option C is wrong because manually adjusting TPM during peak times is reactive, not proactive, and cannot handle sudden spikes instantly due to deployment update delays. Option D is wrong because deploying multiple instances in different regions does not solve quota limits within a single region and introduces latency and complexity without guaranteed scaling within your allocated quota.

Why this answer

How to eliminate wrong answers

Option A is wrong because top_p (nucleus sampling) controls the cumulative probability threshold for token selection, influencing randomness and diversity of output, but it does not specifically penalize repeated phrases. Option B is wrong because presence_penalty penalizes tokens that have appeared at least once in the text, encouraging the model to talk about new topics, but it does not target the frequency of repetition of the same phrases. Option C is wrong because temperature controls the randomness of token selection by scaling the logits before softmax, affecting creativity and variability, but it has no direct mechanism to prevent repetition of phrases.

Why this answer

How to eliminate wrong answers

Option A is wrong because using the global Azure AI services endpoint routes traffic through Microsoft's global network and may process data in any Azure region worldwide, which violates the requirement to keep data within the EU. Option C is wrong because enabling data encryption at rest and in transit protects data confidentiality but does not control the geographic location where data is stored or processed; it does not address data residency. Option E is wrong because storing data in Azure Blob Storage with hot tier is a storage performance and cost choice, not a data residency control; the storage account itself must be deployed in an EU region and configured with appropriate replication settings to meet residency requirements.

Why this answer

How to eliminate wrong answers

Option B is wrong because an incorrect API version would typically result in an HTTP 400 Bad Request or 404 Not Found error, not a successful response with zero entities. Option C is wrong because the document language parameter is optional; if omitted, the API auto-detects the language, and even if set to 'es', Spanish text would still return entities if present. Option D is wrong because an incorrect region endpoint would cause a connection or authentication failure (e.g., 401 Unauthorized or 403 Forbidden), not a valid response with no entities.

Why this answer

How to eliminate wrong answers

Option A is wrong because deleting and recreating the Cognitive Services resource would cause a complete loss of service and all associated configuration, resulting in significant downtime. Option B is wrong because regenerating key1 immediately would invalidate the key currently used by the application, causing authentication failures and downtime until the application is updated with the new key1. Option C is wrong because regenerating both keys at the same time would invalidate all active keys, leaving no valid key for the application to use, causing immediate downtime.

Why this answer

Why this answer

Why this answer

How to eliminate wrong answers

Option A is wrong because the Direct Line channel is designed for custom application integration, not for native Teams distribution, and embedding it in a Teams tab would not enforce the required Entra ID authentication at the channel level. Option B is wrong because the Web channel uses anonymous or generic authentication by default, and sharing a link in Teams does not restrict access to the company's Entra ID tenant. Option C is wrong because turning off authentication on the Teams channel would allow any user, including unauthenticated or external users, to access the agent, violating the security requirement.

Why this answer

Why this answer

Why this answer

How to eliminate wrong answers

Option B is wrong because removing the raiPolicyName property does not affect throughput provisioning; it only removes content filtering or responsible AI policies, which are unrelated to capacity allocation. Option C is wrong because increasing capacity to 200 only scales the number of tokens per minute under the current SKU (Standard), but does not change the SKU to provisioned throughput; PTU requires the SKU name change, not just a higher capacity value. Option D is wrong because changing the model format to 'GPT-4' does not enable provisioned throughput; PTU is a SKU-level setting independent of the model version, and GPT-4 can also be deployed with Standard SKU.

Why this answer

How to eliminate wrong answers

Option A is wrong because Azure AI Document Intelligence is a pre-built skill for document analysis, not a replacement for a custom Azure Function that performs specialized logic; it cannot handle arbitrary custom processing that may cause timeouts. Option C is wrong because Azure Functions have a maximum timeout of 10 minutes (Consumption plan) or 30 minutes (Premium plan), not 24 hours; setting a 24-hour timeout is impossible and would violate platform limits. Option D is wrong because moving the custom skill to Azure AI Language custom entity recognition does not solve timeout issues—it changes the service entirely and still requires a custom component if the logic is not entity recognition, and it does not address the underlying timeout problem.

Why this answer

Why this answer

How to eliminate wrong answers

Option A is wrong because translating all input to English before sending to CLU introduces latency, potential translation errors, and loss of cultural or linguistic nuances, which degrades intent recognition accuracy; the CLU service is designed to handle multiple languages natively without a separate translation step. Option C is wrong because using the Translator service to detect language and route to language-specific CLU endpoints adds unnecessary complexity and overhead; CLU's multi-lingual detection handles language identification and intent recognition in a single model, making separate endpoints redundant. Option D is wrong because creating separate CLU projects for each language duplicates effort, increases maintenance costs, and prevents the model from leveraging cross-lingual transfer learning that improves accuracy for low-resource languages; a single multi-lingual project is the recommended pattern.

Why this answer

How to eliminate wrong answers

Option A is wrong because adding synonyms expands query matching but does not re-rank results based on semantic understanding; it only broadens recall, not precision. Option B is wrong because custom scoring profiles operate on lexical term frequency and field weights, not on the deep neural network models that semantic ranking uses to understand query intent. Option D is wrong because changing the index analyzer affects tokenization and language-specific stemming, not the semantic re-ranking stage that determines which documents are most contextually relevant.

Why this answer

Why this answer

Why this answer

How to eliminate wrong answers

Option A is wrong because Azure AI Vision Image Analysis pre-built models are designed for general-purpose tasks (e.g., describing images, detecting common objects) and cannot be retrained on custom classes like 'defective' vs 'non-defective'. Option C is wrong because object detection identifies and locates multiple objects within an image, which is overkill for a simple binary classification task where only the presence of a defect matters, not its location. Option D is wrong because training a deep learning model from scratch with only 500 images per class would likely result in poor generalization and overfitting, requiring significantly more data and computational resources.

Why this answer

How to eliminate wrong answers

Option B is wrong because increasing the confidence threshold does not improve extraction accuracy; it only filters out lower-confidence results, which may discard valid extractions and reduce recall without addressing the root cause of poor recognition for new formats. Option C is wrong because Azure AI Language service is designed for text analytics (e.g., sentiment, key phrase extraction, language detection) and does not provide document structure analysis or table/field extraction capabilities required for invoice processing. Option D is wrong because increasing the TPS limit only affects throughput and scalability, not the quality or accuracy of extraction results; it does not help the model recognize new vendor formats.

Why this answer

How to eliminate wrong answers

Option A is wrong because training data changes are automatically saved when you edit the dataset in Azure AI Language; the issue is not about saving but about deployment. Option B is wrong because rebuilding the project from scratch is unnecessary; you can retrain and redeploy the same project without recreating it. Option C is wrong because Azure AI Language endpoints do not have a client-side or server-side caching mechanism that would serve stale model predictions; the endpoint simply returns results from whichever model is currently deployed.

Why this answer

How to eliminate wrong answers

Option A is wrong because Azure AI Foundry is a development environment for building and managing AI models, but it does not directly import documents into a question-answering knowledge base; the custom question answering feature is the correct service for this task. Option B is wrong because creating an index in Azure AI Search is used for full-text or vector search, not for the structured question-answer pair extraction and management that custom question answering provides. Option D is wrong because Azure AI Bot Service is a framework for building conversational bots, not a tool for importing documents and creating a knowledge base; the knowledge base must be created first using the custom question answering feature before a bot can consume it.

Why this answer

How to eliminate wrong answers

Option A is wrong because increasing the temperature to 0.9 increases randomness and creativity, which would make the model more likely to hallucinate or deviate from the retrieved documents, worsening the contradiction problem. Option B is wrong because increasing the chunk size to 2000 tokens may include more irrelevant or noisy text per chunk, diluting the precision of the retrieved context and potentially introducing contradictions. Option E is wrong because a system message instructing the model to only use provided context is a soft instruction that the model can ignore or override, especially with high temperature or ambiguous prompts; it does not enforce strict grounding like the parameter in Option C.

Why this answer

Why this answer

How to eliminate wrong answers

Option A is wrong because storing clinical notes in Azure Storage with Azure AD authentication only does not address network-level isolation; the storage account could still be accessible over the public internet, violating HIPAA's requirement to protect ePHI in transit and at rest. Option B is wrong because enabling public network access with a firewall rule to restrict IP addresses still exposes the resource to the public internet, which is not sufficient for HIPAA compliance; private endpoints are the recommended approach for regulated data. Option D is wrong because using a system-assigned managed identity for authentication controls identity but does not provide network isolation; it must be combined with a private endpoint and disabled public access to meet HIPAA requirements.

Why this answer

How to eliminate wrong answers

Option A is wrong because Azure Functions with an HTTP trigger can call external APIs but does not natively support on-premises Windows Authentication or provide a secure bridge to on-premises resources without additional networking components like VPN or gateway. Option B is wrong because Azure VPN Gateway establishes a site-to-site or point-to-site VPN connection to the on-premises network, but it does not handle API-level authentication or credential management for Windows Authentication; it only provides network-level connectivity. Option C is wrong because Azure Logic Apps with a connector can integrate with on-premises systems using the on-premises data gateway, but the standard connectors do not natively support Windows Authentication for custom APIs; the on-premises data gateway is required, and it is typically paired with API Management for secure API exposure.

Why this answer

How to eliminate wrong answers

Option A is wrong because bot-level authentication in Copilot Studio does not support requiring a specific role directly; role-based access must be configured at the topic level after Entra ID authentication is set up. Option C is wrong because enabling audit logging in Microsoft Purview records activities but does not restrict access to topics or enforce authentication. Option D is wrong because data loss prevention policies in Microsoft Purview prevent data exfiltration but do not control who can access specific topics within a Copilot Studio agent.

Why this answer

How to eliminate wrong answers

Option A is wrong because if the flow connection were deleted, the flow would fail even during manual testing, not just when triggered by the agent. Option C is wrong because the user signing in again would not change the identity used by the agent; the agent uses its own service principal, not the user's credentials. Option D is wrong because the CRM connector permissions are already sufficient for the manual test to succeed; the issue is specifically that the agent's identity lacks those permissions, not that the connector itself needs additional grants.

Why this answer

Why this answer

How to eliminate wrong answers

Option A is wrong because `az cognitiveservices account deployment list` returns all deployments in the account, not the status of a specific deployment, and does not directly indicate readiness for inference. Option B is wrong because `az cognitiveservices account deployment delete` removes the deployment entirely, which is destructive and unrelated to checking readiness. Option D is wrong because `az cognitiveservices account deployment create` initiates a new deployment or updates an existing one, but it does not check the current state; it is used to create or modify, not to verify.

Why this answer

How to eliminate wrong answers

Option B is wrong because Azure OpenAI Service with GPT-4 prompt engineering is a generative AI approach that requires extensive prompt tuning and may hallucinate or produce inconsistent extractions, especially for structured fields in legal contracts; it is not optimized for high-accuracy, deterministic extraction from PDFs. Option C is wrong because Azure AI Search with built-in blob indexing provides full-text search and metadata extraction but does not perform custom field-level extraction of clauses, parties, or dates without additional custom skills or AI enrichment pipelines. Option D is wrong because Azure AI Language custom named entity recognition is designed for extracting entities from unstructured text (e.g., news articles, social media) and does not natively handle PDF layout, tables, or multi-page document structure, requiring significant preprocessing and labeling for contract-specific fields.

Why this answer

How to eliminate wrong answers

Option A is wrong because Custom Translator is a feature for building custom translation models tailored to specific domain terminology, not for preserving document formatting or layout. Option C is wrong because Azure AI Document Intelligence (formerly Form Recognizer) is used for extracting text, key-value pairs, and tables from documents, not for translating them. Option D is wrong because Text Translation only handles plain text strings and cannot preserve the formatting or layout of an entire document.

Why this answer

How to eliminate wrong answers

Option A is wrong because changing the 'severityThreshold' for 'SelfHarm' to 'low' would affect the SelfHarm category, not the Hate category, and would block more SelfHarm content than needed, failing to address the requirement. Option B is wrong because 'contentTypes' typically refer to modalities like text, image, or code, not severity levels; restricting to only 'Normal' would block all non-normal content types, not specifically Hate content at medium severity. Option D is wrong because 'policyAction' with 'block' is not a valid property in the Azure OpenAI content filtering JSON policy; the blocking behavior is controlled by the 'severityThreshold' and the category definitions, not by an explicit 'block' action.

Why this answer

How to eliminate wrong answers

Option B is wrong because fine-tuning a model does not generate data locally; it adapts a pre-trained model to a specific task, and the data generation still requires API calls or local inference, which is not cost-effective for large volumes. Option C is wrong because the streaming API is designed for real-time, low-latency responses, not for high-throughput batch processing, and managing multiple concurrent connections increases complexity and cost without the efficiency of batching. Option D is wrong because deploying a model on Azure Functions and calling it in parallel introduces overhead from serverless scaling and per-execution costs, which is less cost-effective than the batch API's optimized queuing and processing.

Why this answer

How to eliminate wrong answers

Option A is wrong because scoring profiles influence the ranking of search results based on fields, functions, or weights, but they do not alter how text is tokenized or stemmed for different languages; they cannot ensure cross-lingual retrieval relevance. Option C is wrong because semantic search improves result relevance by understanding query intent and context using deep learning models, but it does not provide language-specific tokenization or stemming; it works on top of existing analyzers and is not a substitute for language analyzers. Option D is wrong because synonym maps expand queries with equivalent terms but do not handle language-specific linguistic rules like stemming or diacritic normalization; they are language-agnostic and cannot adapt to different languages' morphological structures.

Why this answer

How to eliminate wrong answers

Option A is wrong because configuring variables captures and stores user input after it has been processed, but does not perform intent recognition or map utterances to topics. Option B is wrong because actions (such as calling Power Automate flows or APIs) are used to execute logic after a topic is triggered, not to understand the user's intent from free-text input. Option C is wrong because custom entities extract specific data points (like product names or dates) from utterances, but they do not map the entire utterance to a topic; entities are used within a topic to refine understanding, not to trigger the topic itself.

Why this order

Why this answer

Why this answer

How to eliminate wrong answers

Option A is wrong because QnA Maker is designed for FAQ-style question answering with predefined QnA pairs, not for handling multiple intents in a single utterance; it lacks intent recognition and orchestration capabilities. Option B is wrong because a single CLU project with multiple intents can only classify one intent per utterance at a time, not handle multiple intents in a single utterance; it does not support splitting or routing the utterance to different services. Option D is wrong because Azure Bot Service with multiple dialogs manages conversation flow and state, but it does not natively handle multiple intents in a single utterance; it relies on an underlying NLP service like CLU or LUIS for intent recognition, and the orchestration of multiple intents must be implemented externally.

Why this answer

How to eliminate wrong answers

Option A is wrong because using the push API exclusively requires building a custom application to poll all sources, which is unnecessary overhead for SQL and SharePoint when native indexers exist, and it does not leverage Azure AI Search's built-in change tracking and scheduling capabilities. Option B is wrong because a single indexer cannot read from multiple heterogeneous data sources; each indexer is bound to one data source definition, and you must create separate indexers for SQL, SharePoint, and REST APIs. Option C is wrong because Azure AI Search does not support creating custom indexers for REST APIs; the only way to index data from a custom REST API is via the push API, not an indexer.

Why this answer

Why this answer

How to eliminate wrong answers

Option A is wrong because disabling network ACLs entirely would remove all IP-based restrictions but would also require changing the defaultAction to 'Allow', which is an overly permissive approach that violates security best practices and is not necessary when only specific IPs need access. Option C is wrong because changing defaultAction to 'Allow' and removing all rules would open the service to the entire internet, which is a severe security risk and not required when you can simply add the developers' IPs. Option D is wrong because a virtual network rule requires the developers' machines to be connected to the specified virtual network (typically via VPN or ExpressRoute), and since they are not connected to the corporate VPN, a virtual network rule cannot be applied to their local machines.

Why this answer

How to eliminate wrong answers

Option B is wrong because customer-managed keys add complexity to the deployment process (you must create and manage a Key Vault, set permissions, and configure key rotation), not simplify it. Option C is wrong because encryption keys have no impact on model inference speed or accuracy; performance is determined by model size, token limits, and compute resources. Option D is wrong because CMK typically increases operational costs due to the need for additional Key Vault resources, key management overhead, and potential charges for key operations.

Why this answer

How to eliminate wrong answers

Option A is wrong because rate limiting controls the number of requests processed per time period to manage resource usage and prevent abuse, but it has no mechanism to detect or remove PII from summaries. Option C is wrong because while a system message can instruct the model to exclude PII, it relies entirely on the model's compliance and can be bypassed by adversarial prompts or model hallucinations, offering no guaranteed enforcement. Option D is wrong because the 'Add your data' feature grounds the model on your own documents for retrieval-augmented generation, but it does not include any PII detection or redaction capability—it simply provides additional context without filtering sensitive content.

Why this answer

How to eliminate wrong answers

Option A is wrong because a timer trigger runs on a fixed schedule (e.g., every 5 minutes), which introduces latency and cannot react immediately to uploads; it would require polling the container for new blobs. Option C is wrong because Azure Batch is designed for large-scale parallel compute jobs with job schedules, not for real-time event-driven triggers on individual blob uploads. Option D is wrong because a recurrence trigger in Logic Apps also runs on a schedule, not event-driven, and would similarly require polling, missing the immediate processing requirement.

Why this answer

How to eliminate wrong answers

Option B (Azure Monitor) is wrong because it focuses on performance metrics, logs, and alerts for application health and resource utilization, not on cost tracking or billing data. Option C (Azure Service Health) is wrong because it monitors service-level issues, outages, and planned maintenance across Azure services, not usage or cost metrics. Option D (Azure Advisor) is wrong because it provides best-practice recommendations for optimizing cost, performance, and reliability, but it does not directly monitor or report on actual usage and costs in real time.

Why this answer

Why this answer

Why this answer

How to eliminate wrong answers

Option A is wrong because Function calling is designed for structured API interactions (e.g., calling external services or databases) and does not directly process uploaded file contents or execute arbitrary code. Option B is wrong because KQL (Kusto Query Language) is used for querying Azure Data Explorer and log analytics data, not for document extraction or summarization. Option D is wrong because a Knowledge base stores pre-indexed information for retrieval-augmented generation (RAG) but does not dynamically execute code to process newly uploaded documents or extract content on the fly.

Why this answer

How to eliminate wrong answers

Option A is wrong because adding additional predefined models does not improve accuracy for a specific field; predefined models are fixed and cannot be retrained or customized for your data. Option B is wrong because OCR enhancement improves text recognition quality but does not address the model's ability to correctly interpret and extract the 'total amount' field from the recognized text. Option C is wrong because increasing the confidence threshold only filters out low-confidence predictions, it does not improve the underlying model's extraction accuracy; it may reduce false positives but will not correct mis-extractions.

Why this answer

Why this answer

How to eliminate wrong answers

Option A is wrong because the prebuilt-layout model is designed for extracting text and structure (like tables and selection marks), not for custom clause extraction from legal contracts, and it would not leverage the firm's existing labeled data. Option B is wrong because increasing the OCR confidence threshold only filters out low-confidence text recognition results; it does not improve the model's ability to correctly classify or extract clauses from a different font and layout. Option C is wrong because labeling 15 more contracts with the original layout and retraining the model would not address the new layout variation; it would only reinforce the existing model's performance on the original layout, wasting effort and cost.

Why this answer

How to eliminate wrong answers

Option B is wrong because disabling retry attempts entirely would cause the system to fail on any transient error, making it less resilient and increasing the likelihood of incomplete agent tasks. Option C is wrong because increasing the global timeout for all agents does not address the root cause of transient API failures; it only delays the timeout, potentially masking underlying issues and wasting resources. Option D is wrong because switching to synchronous agent calls would block agent execution, reducing concurrency and potentially worsening timeout issues by making the system less responsive to failures.

Why this answer

How to eliminate wrong answers

Option A is wrong because Whisper is a speech-to-text model used for transcribing audio, not for generating images. Option C is wrong because GPT-4 is a large language model focused on text generation and understanding, lacking native image generation capabilities. Option D is wrong because Codex is a model specialized in code generation and completion, not image creation.

Why this answer

How to eliminate wrong answers

Option A is wrong because local deployment (e.g., a local Docker container or Jupyter notebook) is intended for development and testing only, not for production-grade low-latency real-time inference, as it lacks scalability, load balancing, and network-level optimizations. Option B is wrong because Azure Batch is designed for high-throughput, parallel batch processing jobs (e.g., offline scoring of large datasets) and is not optimized for low-latency real-time inference due to its job-queue scheduling overhead and lack of persistent endpoints. Option C is wrong because Azure Functions, while serverless and capable of handling HTTP triggers, has a cold-start latency problem (often 1-10 seconds) and limited GPU support, making it unsuitable for sub-second real-time inference workloads.

Why this answer

Why this answer

Why this answer

How to eliminate wrong answers

Option A is wrong because the OCR API is a legacy endpoint that supports fewer languages and is not optimized for document-like images such as business cards; it is primarily intended for simple, single-language text extraction from natural scenes. Option B is wrong because Recognize Text is a deprecated operation in the Computer Vision API that has been replaced by the Read API; using it would not be a best practice and may lack support for multiple languages and modern features. Option D is wrong because Describe Image generates a human-readable caption describing the image content, not extracting text, so it cannot fulfill the requirement of extracting text from business cards.

Why these pairings

Why this answer

Why this answer

How to eliminate wrong answers

Option B is wrong because customer-managed keys (CMK) only control encryption at rest and do not prevent data from being used for model improvement or abuse monitoring; CMK is about key ownership, not data governance for training. Option C is wrong because a private endpoint restricts network access to the service but does not affect how Microsoft processes or stores data for model improvement; it only secures the connection path, not the data usage policy. Option D is wrong because data residency configuration ensures data is stored in a specific geographic region but does not prevent Microsoft from using that data for abuse monitoring or base model training; residency is about location, not usage rights.

Why this answer

How to eliminate wrong answers

Option A is wrong because Azure AI Computer Vision is optimized for general image analysis (e.g., object detection, OCR for raw text) but lacks the specialized prebuilt models for extracting line-item tables from invoices with varied layouts. Option B is wrong because Azure AI Language Service focuses on text analytics (e.g., sentiment, key phrase extraction, NER) and is not designed for document structure understanding or table extraction from PDFs. Option C is wrong because Azure AI Search is a search indexing and query service, not a document extraction tool; it can index extracted data but cannot perform the initial extraction of line-item details from invoices.

Why this answer

How to eliminate wrong answers

Option A is wrong because reducing image size can improve inference speed but does not enable offline inference—the app still requires network access to the prediction endpoint. Option C is wrong because increasing the prediction API timeout only extends the wait time for a response, which does not solve the underlying network latency or enable offline operation. Option D is wrong because deploying to an Azure IoT Edge device moves inference to an edge gateway, not the mobile device itself, and still requires network connectivity between the mobile app and the edge device, failing to provide true offline inference on the mobile device.

Why this answer

How to eliminate wrong answers

Option A is wrong because creating a Cognitive Services multi-service account does not increase throughput for Document Intelligence; it simply combines multiple AI services under a single endpoint and key, and each service still has its own tier limits. Option C is wrong because the Free tier (F0) is capped at 20 transactions per minute (1,200 per hour) and has a maximum of 5 TPS, which cannot reliably handle 1,000 invoices per hour with 5-second latency, and retry policies do not overcome throughput limits. Option D is wrong because deploying Document Intelligence resources in multiple regions does not increase throughput for a single workload; it is used for geo-redundancy or data residency, not for scaling capacity within a single region.

Why this answer

How to eliminate wrong answers

Option A is wrong because a low `max_tokens` value would cause the response to stop at the token limit, but the response would not necessarily be cut off mid-sentence; it would simply end after the specified number of tokens, which could be at a natural break. Option B is wrong because a high `temperature` setting increases randomness and creativity but does not cause early termination; it affects the probability distribution of token selection, not the stopping condition. Option D is wrong because the system message provides context or instructions for the model's behavior, but its absence would not cause a mid-sentence cut-off; it might lead to less relevant or less structured responses, but the response would still complete naturally unless another parameter stops it.

Why this answer

How to eliminate wrong answers

Option B is wrong because Azure AI Search custom skills are used to enrich search indexes, not to extract fields from documents during ingestion; they operate on already-extracted content, not on raw scanned invoices. Option C is wrong because training a custom NER model in Azure AI Language requires a labeled dataset and is overkill for a fixed regex pattern; it is designed for entity recognition in text, not for structured extraction from scanned documents. Option D is wrong because while GPT-4 with vision can extract codes, it introduces unnecessary cost, latency, and complexity compared to a native regex-based extraction in Document Intelligence, which is purpose-built for this task.

Why this answer

Why this answer

How to eliminate wrong answers

Option A is wrong because Text Analytics for Health is designed for medical entities (e.g., diagnoses, medications) and does not support generic order numbers or custom categories, leading to poor accuracy for support tickets. Option C is wrong because the Prebuilt Entity Extraction skill in Azure AI Search is a limited set of generic entities (e.g., person, location) and cannot be trained for domain-specific categories like issue types or order number patterns. Option D is wrong because Conversational Language Understanding (CLU) focuses on intent and entity extraction for conversational flows, which is overkill for simple entity extraction and introduces unnecessary complexity in training and deployment.

Why this answer

How to eliminate wrong answers

Option A is wrong because Azure Custom Vision is a service for training custom image classification and object detection models on images, not for analyzing pre-recorded videos with timestamp generation. Option B is wrong because Azure Form Recognizer is designed to extract text and structure from documents (e.g., invoices, forms), not for video analysis or object detection. Option C is wrong because Azure Computer Vision provides image analysis APIs (e.g., object detection in static images) but lacks native video processing capabilities and timestamp generation; it would require additional custom logic to handle video frames sequentially.

Why this answer

Why this answer

How to eliminate wrong answers

Option B is wrong because Azure AI Vision OCR only extracts raw text from images without understanding document structure or field semantics, and Azure AI Language provides NLP capabilities but cannot classify documents by layout or extract structured invoice fields. Option C is wrong because Azure AI Form Recognizer is deprecated and replaced by Azure AI Document Intelligence; using the deprecated service is not recommended for new solutions. Option D is wrong because the pre-built invoice model can extract fields like total amount and date, but it cannot classify invoices by vendor—it assumes all documents are invoices and does not support custom classification logic.