Microsoft AI-103 Practice Test 5
Question 40
A regulatory PDF has a two-column layout, and extracted text mixes unrelated sentences from each column. What setting should be evaluated?
- Use layout-aware reading-order reconstruction before chunk creation.
- Sort all OCR text blocks by their top-left vertical coordinate only.
- Remove column coordinates and join extracted text into one unstructured string.
- Increase vector-index dimensions while keeping the interleaved source text.
Explanation
Correct Answer
A:Column-aware processing preserves intended sentence order; plain coordinate sorting can interleave independent passages.
Incorrect Answers
B:Global top-to-bottom sorting can interleave lines from the left and right columns incorrectly.
C:Discarding layout geometry makes the reading-order ambiguity harder to detect and repair.
D:Embedding size cannot restore the correct reading order lost during text extraction.
Question 41
A model responds with a usage object showing prompt and completion tokens. What should the app record?
- Log only the model's response text, discarding prompt-token counts, completion-token counts and serving deployment identity.
- Capture usage counts with the deployment name for per-model cost attribution
- Use retry-after-aware backoff but record no deployment ID, prompt token count or generated-token count.
- Use the correct service endpoint and API contract for the chosen model deployment
Explanation
Correct Answer
B:Usage telemetry tied to deployment identifiers supports spend and optimization decisions. Here the relevant capability must let the team account for input and output model tokens. An alternative targeting supply image evidence in model input solves another problem.
Incorrect Answers
A:Text alone cannot attribute usage or cost to the deployed model. Store the API's token usage fields with model and deployment identifiers for correct accounting.
C:Backoff reduces throttling but does not attribute inference consumption. Record input and output token usage together with model deployment and request identifiers for accurate per-model cost analysis.
D:Foundry project management routes and model inference endpoints are not necessarily interchangeable. Here the needed outcome is to account for input and output model tokens; an option designed to select the correct model serving API would not directly provide it.
Question 42
A customer helpdesk is selecting models for an offline fixed-label kiosk and a cloud-based image-plus-text complaint workflow. Which TWO model decisions are justified by those distinct requirements? Choose TWO.
- Deploy a compact language model suitable for local inference
- Use one text-only remote classifier for both workflows and ask kiosk staff to reconnect during outages or transcribe visual damage manually.
- Select a multimodal model that accepts images together with text
- Use a cloud-only large model for the offline kiosk even though its required classifier must operate during network outages.
- Select one text-only model from a medical-abbreviation leaderboard for both the offline kiosk and the image-plus-text workflow.
Explanation
Correct Answers
A:A small model can meet bounded classification needs with a smaller memory and latency footprint, while local execution avoids cloud connectivity dependencies. The kiosk needs a small locally runnable model, whereas image-grounded complaints require an image-capable multimodal model; neither substitutes for the other.
C:A text-only model cannot directly inspect pixels; a multimodal input model can jointly use the submitted image and the accompanying question.
Incorrect Answers
B:The kiosk must work offline and the complaint workflow must process image pixels. A cloud-only text model fails both constraints; each workload needs its own validated capability.
D:A cloud-only model cannot satisfy a requirement to classify locally when connectivity is unavailable. Choose a compact deployable local model and validate its fixed-label accuracy and hardware footprint.
E:Text-only medical results do not show offline hardware suitability or image understanding. The kiosk and photo-complaint application have distinct requirements that need different tested model capabilities.
Question 43
A media application must create moving assets in approved dimensions and preserve how they were generated. Which THREE controls meet those different requirements? Choose THREE.
- Select an available text-to-video generation capability and suitable output constraints
- Store generation request IDs, prompt versions, input references and output hashes
- Use an image-only generator to produce a single frame and label it as a full video while omitting generation request provenance.
- Cache a single still frame as the delivered video and retain no generation request or source provenance.
- Apply supported video duration, aspect-ratio and resolution controls where available
- Generate an image from a supported text prompt using an image-capable model
Explanation
Correct Answers
A:Video generation produces a temporal sequence; a still-image API will not produce moving output on its own.
B:Provenance supports reproducibility, moderation review and asset management.
E:Generation controls should be chosen from the model's actual supported capabilities and validated against delivery constraints.
Incorrect Answers
C:A still image cannot satisfy the required moving-media output, and omitting provenance undermines review. Video generation, supported output dimensions and traceable request data are separate requirements.
D:Caching a still image neither creates the required moving output nor establishes how a video was generated. Use supported video generation plus versioned request and source records.
F:Text-to-image synthesis creates novel visual content from a natural-language description. That mechanism concerns create a synthetic image from prose instructions, whereas this case tests how to generate video, constrain format, and track origins.
Question 44
A medical study summary omits adverse events because they occur in an appendix. What retrieval strategy helps?
- Summarize only the article abstract and main body while excluding the adverse-event appendix
- Summarize only the main article body and exclude the appendix from retrieval because its position suggests lower priority.
- Retrieve relevant sections across the whole source before summarization
- Redact or mask sensitive fields before publishing the summary
Explanation
Correct Answer
C:A summary cannot include critical evidence that prompt assembly never retrieved. The decisive difference is whether the design can include important evidence located outside opening sections; the competing function serves to summarize a long contract with traceable evidence.
Incorrect Answers
A:Excluding an appendix known to contain safety outcomes silently biases the summary. Retrieval must cover relevant source sections, including attached evidence beyond the main narrative.
B:Adverse events may appear in appendices and remain material to the study. Retrieve relevant evidence across the complete source, preserving section and page references.
D:Data minimization must apply to summarization outputs as well as original inputs. That is relevant to remove personal data from a shareable summary, but the present scenario requires an ability to include important evidence located outside opening sections.
Question 45
An application treats low-confidence OCR of an account number as verified. What must be changed?
- Accept the highest OCR hypothesis for an account number without format or confidence review
- Mark every OCR-extracted account number as verified whenever the parser returns syntactically valid digits, regardless of confidence.
- Validate extracted identifiers and route uncertain OCR outputs for review
- Apply optical character recognition to scanned image pages
Explanation
Correct Answer
C:Uncertain recognition should not silently become authoritative structured data. Here the relevant capability must let the team avoid accepting incorrect low-confidence OCR. An alternative targeting retain structured tabular data from images solves another problem.
Incorrect Answers
A:A top-ranked recognition guess can still be wrong, especially for high-impact identifiers. Format validation and an uncertainty escalation path are needed before account data is treated as verified.
B:A value can be numerically well formed yet misread from the scan. Validate the identifier against domain rules and send low-confidence extraction for review before using it.
D:OCR converts visual text into searchable characters when no text layer is available. Here the needed outcome is to avoid accepting incorrect low-confidence OCR; an option designed to extract text from scanned pages would not directly provide it.
Question 46
A user asks about a particular customer case, but the search index contains records from several customers. Where must tenant isolation be applied?
- Apply tenant and user authorization filters within the retrieval query and enforce them at the data source.
- Retrieve all customer documents and let the model decide what to reveal.
- Apply only a string-matching filter to the answer after generation.
- Use one service-level API key with unrestricted cross-customer index reads.
Explanation
Correct Answer
A:Filtering only the final generated answer risks leaking unauthorized retrieved text into the model context.
Incorrect Answers
B:Unauthorized records enter model context before any response-level censorship and violate tenant isolation.
C:Post-generation text filtering cannot undo exposure of restricted records during retrieval and prompting.
D:A shared broad key does not enforce the requesting user's authorization to individual records.
Question 47
A policy assistant must answer employees using an indexed set of changing internal handbooks. Which component should provide grounded retrieval?
- Feed the entire handbook collection into one long model prompt and skip building a searchable, versioned knowledge index.
- Combine conversation state with a controlled knowledge retrieval connection
- Expose a write-capable employee-record API instead of connecting the policy agent to a searchable handbook index.
- Use Azure AI Search as a searchable knowledge index for grounded queries
Explanation
Correct Answer
D:Azure AI Search can index approved handbook text and expose searchable, citeable passages to a grounded generation workflow. Keep document versions, retrieval permissions and effective dates synchronized; merely attaching unindexed files does not establish reliable source retrieval.
Incorrect Answers
A:A large static prompt does not provide scalable, current, access-controlled document retrieval. Index authoritative handbook passages in a searchable store and retrieve them as needed.
B:Continuity is supplied by retained state and updated facts by knowledge retrieval; the application should govern both. Here the needed outcome is to retrieve relevant indexed company passages; an option designed to maintain agent continuity with grounded knowledge would not directly provide it.
C:An API for employee-record changes cannot retrieve authoritative policy passages from the indexed handbooks. A grounded policy assistant needs governed retrieval with source evidence, not an unrelated mutation tool.
Question 48
A photographer wants to replace a distracting sign in one small region while leaving the rest of the image unchanged. Which operation fits?
- Apply a global image style change to remove the sign without protecting the rest of the photograph.
- Use mask-based image inpainting on the marked region
- Use a multi-frame video edit to replace a small sign in a single still photograph.
- Run a full-image generative redraw and compare only the replaced sign, accepting changes to nearby labels and product details.
Explanation
Correct Answer
B:Inpainting targets a defined spatial area instead of regenerating an entire image unnecessarily. The decisive difference is whether the design can replace only a selected image region; the competing function serves to reproduce a localized image edit.
Incorrect Answers
A:A whole-image style operation can alter details outside the sign. The requested local replacement calls for a marked edit region with the surrounding pixels preserved.
C:Temporal video editing manages changes across frames; this task concerns one region of a still image. Use supported image inpainting with a mask protecting unaffected pixels.
D:Unrestricted regeneration may modify regions the photographer wanted preserved. A masked inpainting operation targets the sign while protecting surrounding content.
Question 49
A summarizer copies personal identifiers into a public project digest. What pipeline step should be added?
- Use source citations alone to justify publishing raw personal identifiers in the digest
- Redact or mask sensitive fields before publishing the summary
- Publish the digest with all original identifiers intact and rely on a later keyword search to find names after readers have seen them.
- Retrieve relevant sections across the whole source before summarization
Explanation
Correct Answer
B:Data minimization must apply to summarization outputs as well as original inputs. In this scenario the team must remove personal data from a shareable summary. A feature used to summarize a long contract with traceable evidence does not substitute for it.
Incorrect Answers
A:Traceable citations do not make personal identifiers suitable for public distribution. The dissemination path needs data minimization and masking before a public digest is released.
C:Postpublication discovery does not prevent disclosure. Redact or mask sensitive fields in the summary before it is released to the public audience.
D:A summary cannot include critical evidence that prompt assembly never retrieved. This helps when a team needs to include important evidence located outside opening sections; it does not establish how to remove personal data from a shareable summary.
Question 50
An OCR system confidently transcribes a machine serial number, but the value corresponds to a different device in the authoritative asset registry. Which check should run before using it for a warranty lookup?
- Accept high-confidence OCR identifiers as authoritative even when the equipment registry disagrees.
- Cross-check the recognized serial with the authoritative asset registry and flag any model or device mismatch for investigation.
- Accept the recognized serial because the OCR confidence score is high and omit comparison with the device identified by the asset registry.
- Use layout-aware reading order and column segmentation
Explanation
Correct Answer
B:Optical character recognition can be confidently wrong or applied to the wrong device. The recognized identifier should be checked against an independent asset register, and conflicts reviewed before authorizing a warranty operation. This differs from checking low-confidence characters or OCR layout quality. A high OCR confidence score does not prove the value belongs to the claimed device; a trusted registry supplies independent identity validation.
Incorrect Answers
A:Recognition confidence measures a model output, not real-world ownership or identity consistency; conflicting registry evidence must not be ignored.
C:OCR confidence measures recognition, not whether the value belongs to this machine. Validate the serial against the authoritative inventory before initiating a warranty lookup.
D:Page layout structure is necessary for sensible text order in multi-column documents. This does not provide the required ability to corroborate a confidently recognized identifier against a separate authoritative registry.