AI-103 Practice Questions: Implement computer vision solutions Domain

Correct answer (B): This is a structured business-document extraction scenario, not a generic image-description task. Azure AI Document Intelligence is the best choice because it is designed for layout-aware extraction of fields, key-value pairs, and tables from documents such as invoices. OCR alone can read text, but it does not provide the same document-aware extraction behavior. Captioning and open-ended multimodal chat are less suitable for repeatable, production-grade invoice processing.

Why the other options are wrong:
- Option A: OCR can read text, but the requirement includes structured extraction of fields and tables. Document Intelligence is more appropriate for layout-aware business document processing.
- Option C: Captioning gives a high-level description, not precise extraction of invoice fields and tables.
- Option D: A multimodal chat model can discuss the document, but the requirement is repeatable structured extraction for downstream processing, which is better handled by Document Intelligence.

Correct answer (C): When Azure services support identity-based access, the preferred production design is to use managed identity with RBAC instead of storing service keys in code or configuration. This reduces secret-management risk, supports least privilege, and follows Azure-native security practices. Rotating keys is better than leaving them static, but it is still not the best option when managed identity is available.

Why the other options are wrong:
- Option A: Storing secrets in source control increases exposure risk and is not an acceptable production security practice.
- Option B: Key rotation reduces some risk, but it is still weaker than managed identity and RBAC when the services support identity-based access.
- Option D: Prompts are not a secure place for credentials, and the model should not receive embedded secrets to access resources.

Correct answer (A): The problem described is poor source-image quality. Glare, blur, and skew can significantly reduce OCR accuracy, so the best remediation is to improve capture conditions and handle low-quality images before or during OCR. Simply switching to a larger model is not a reliable fix when the visible text is degraded in the source image.

Why the other options are wrong:
- Option B: Image generation is not the right tool for accurately reading operational label photos.
- Option C: A larger model does not reliably solve poor input quality. The root issue is the source image, not just model size.
- Option D: Keyword tags do not preserve the exact text needed for serial numbers and labels, so this would not meet the OCR requirement.

Correct answer (C): The requirement is not just raw text extraction. The app must identify business fields and semantic relationships across variable layouts, which calls for document extraction or content-understanding capabilities beyond plain OCR.

Why the other options are wrong:
- Option A: Incorrect. OCR alone does not reliably infer higher-level document structure, business fields, or semantic relationships, especially across variable layouts and noisy scans.
- Option B: Incorrect. A caption is not a dependable structured extraction method for invoice fields such as totals and due dates.
- Option D: Incorrect. Tags may classify a receipt, but they do not return the required structured fields.

Correct answer (A): The requirement is to describe existing images, so the workflow should use image understanding rather than image generation. Because the output is public-facing and accessibility-related, a human review step is an appropriate control since automatically generated captions or alt text can still be incomplete or inaccurate.

Why the other options are wrong:
- Option B: Incorrect. Image generation is for creating or editing images from prompts, not for analyzing an existing image and producing reliable alt text.
- Option C: Incorrect. OCR extracts visible text only. Alt text usually requires a broader description of the image, not just words appearing inside it.
- Option D: Incorrect. Tags can support search, but they do not replace descriptive alt text, and this option also ignores the required review step.

Correct answer (B): Successful API responses do not prove the outputs are high quality or safe. Production observability for vision apps should include tracing, latency, failures, and safety-related events, while quality evaluation must be planned explicitly as a separate concern.

Why the other options are wrong:
- Option A: Incorrect. Uptime metrics alone do not measure latency problems, safety issues, or output quality.
- Option C: Incorrect. Removing tracing weakens observability, and occasional surveys do not replace continuous operational metrics or explicit quality evaluation.
- Option D: Incorrect. Watermarking is a governance mechanism for generated content, not a monitoring strategy for latency, failures, safety events, or caption quality.

Correct answer (B): The requirement is periodic detection of a simple visible condition, not end-to-end understanding of the entire video. Sampling frames or key frames and analyzing them as images is the most cost-effective design. Full video reasoning adds unnecessary cost and complexity, OCR is for text, and a text-only model cannot directly interpret camera frames.

Why the other options are wrong:
- Option A: A full video reasoning workflow could work, but it is not the best fit when only periodic visible-state checks are required.
- Option C: OCR is for text extraction, not for determining whether a warning light is visibly on.
- Option D: A text-only model trained on notes does not process images or video frames, so it cannot solve this visual detection task.

Correct answer (A): Because the requirement is simple periodic inspection rather than motion or event understanding over time, sampling frames is sufficient and cost-effective. A design that analyzes every frame is unnecessary here.

Why the other options are wrong:
- Option B: Incorrect. Isolated frame analysis can be sufficient and more cost-effective when the use case is simple periodic inspection rather than motion over time.
- Option C: Incorrect. Image generation does not solve the problem of analyzing existing video content for stock presence.
- Option D: Incorrect. OCR is aimed at text extraction, not general shelf-stock analysis.

Correct answer (C): Multimodal production monitoring should cover operational health and AI-specific behavior together. The best approach is to collect request traces, latency details, errors, safety-related events, and usage telemetry so the team can diagnose performance regressions, rising cost, and unsafe outputs. Narrow metrics such as status codes or image counts do not provide enough evidence to troubleshoot a multimodal pipeline effectively.

Why the other options are wrong:
- Option A: HTTP status codes show failures, but they do not explain latency spikes, higher costs, or unsafe responses.
- Option B: Safety telemetry matters, but it is only one part of observability. The team also needs performance, error, and usage data.
- Option D: Image count alone is too coarse. It does not reveal latency breakdowns, prompt-size effects, tool-call behavior, or safety events.

Correct answer (C): When a multimodal agent can take real-world actions through tools, the secure design is to limit what the agent is allowed to call, apply least-privilege permissions, and add approval gates for high-risk operations. This reduces the impact of prompt injection, embedded malicious instructions, or incorrect model output. Content filters can help, but they do not replace tool restrictions and human approval for risky actions.

Why the other options are wrong:
- Option A: High confidence is not an authorization mechanism. A model can still be wrong or manipulated while appearing confident.
- Option B: Content filters help, but they do not prevent all prompt injection or replace tool-level permission boundaries and approval gates.
- Option D: Trace logging supports investigation and governance. Disabling it weakens observability and does not reduce the actual risk.