Microsoft AI-103 Practice Test 6
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.
Question 51
A new Python AI application must call the correct deployed model, generate machine-readable records, and survive rate-limit bursts. Which THREE implementation choices are necessary? Choose THREE.
- Use bounded concurrency and retry-after-aware backoff for model requests
- Apply a supported JSON schema or structured-output response contract
- Increase the output-token ceiling without adding structured-result validation, authorized model invocation or rate-limit handling.
- Call the deployed model through the supported SDK or inference endpoint using its deployment identifier
- Capture usage counts with the deployment name for per-model cost attribution
- Call only the Foundry project-listing endpoint and treat the returned resource list as an inference response.
Explanation
Correct Answers
A:Client-side admission and backoff prevent synchronized floods during provider throttling. Correct inference routing, structured-output contracts and bounded retries/concurrency address separate runtime needs.
B:Structured output constraints help downstream parsers consume known required properties and types.
D:A deployed model must be addressed through its serving API; a project name alone does not produce an inference.
Incorrect Answers
C:More output tokens do not replace the three required capabilities: a valid serving call, a structured response contract and bounded retry-aware request concurrency.
E:Usage telemetry tied to deployment identifiers supports spend and optimization decisions. Here, however, the stated objective is to invoke model, constrain data format, and handle throttling; this option focuses on account for input and output model tokens.
F:A project enumeration response is not a model completion. A valid model inference route and deployment ID are required.
Question 52
An AI chat application has a predictable constant load and needs reserved model processing capacity instead of competing for shared throughput. Which hosting decision should be evaluated?
- Evaluate a provisioned-throughput deployment and its capacity commitment for the chosen model.
- Use only an ordinary pay-as-you-go deployment with no capacity planning.
- Increase the number of web frontend instances and assume model inference is reserved.
- Provision one larger Azure AI Search service instead of the model capacity.
Explanation
Correct Answer
A:Provisioned throughput allocates model serving capacity for workloads with known utilization; it trades reservation and capacity planning against consumption pricing.
Incorrect Answers
B:Consumption throughput may be suitable for variable demand but does not reserve committed model processing capacity.
C:Web scaling does not allocate guaranteed inference capacity for the selected model deployment.
D:Search-service capacity is independent of provisioned throughput for model inference.
Question 53
A marketing team generates short videos in several cuts, but the subject's clothing changes between adjacent frames. What quality review is necessary?
- Inspect one attractive frame from each clip and ignore transitions.
- Measure video encoding bitrate as the sole continuity acceptance criterion.
- Evaluate temporal consistency across frames and regenerate or edit inconsistent segments.
- Increase speech synthesis quality without rechecking visual frames.
Explanation
Correct Answer
C:Frame-by-frame visual appeal does not guarantee continuity of subject appearance during motion.
Incorrect Answers
A:Isolated frames may look plausible despite inconsistent clothing between adjacent frames and cuts.
B:Bitrate describes compression rather than whether the subject's appearance is consistent.
D:Improved audio does not address visual identity and clothing continuity across generated frames.
Question 54
A document-processing pipeline rejects model output because expected dates are emitted as informal phrases. What control is needed?
- Constrain dates to a documented machine-readable schema and validate them
- Verify that the JSON contains a nonempty date string while accepting phrases such as 'next Thursday' in a field requiring ISO dates.
- Use one versioned schema for every response and test representative edge cases
- Constrain predictions to an explicit allowed enumeration and reject out-of-range values
Explanation
Correct Answer
A:A structured date format enables downstream parsers to reliably interpret extracted values. This resolves the specific need to enforce typed machine-readable date fields, while a measure that helps enforce required structured output fields addresses a different requirement.
Incorrect Answers
B:A present string is not necessarily machine-readable. Apply a typed date representation and validate the actual format and calendar value before accepting the record.
C:A stable typed response contract limits format drift across similar inputs. This helps when a team needs to keep extraction fields consistent across inputs; it does not establish how to enforce typed machine-readable date fields.
D:Fixed-label extraction needs an enumerated schema, not open-ended descriptive prose. The function serves to limit classification to the approved labels, whereas the actual requirement is to enforce typed machine-readable date fields.
Question 55
An archive needs spoken interviews searchable by subject. Which transformation supports that?
- Transcribe audio and index timestamped text segments
- Label the audio file with a general interview topic but never transcribe the speech or retain time offsets.
- Enforce document-level access rules on transcript and audio source references
- Record source recording ID and time offsets for each transcript chunk
Explanation
Correct Answer
A:Transcripts make spoken content text-searchable while timestamps preserve links to the recording. The decisive difference is whether the design can create searchable audio transcripts; the competing function serves to separate speakers in indexed audio.
Incorrect Answers
B:A coarse label is not a searchable record of what speakers said. Transcribed time-stamped utterances support topic retrieval and navigation to relevant moments.
C:Audio-derived transcripts can contain sensitive information and require source-equivalent permissions. Its primary purpose is to apply source ACLs to audio transcripts; the stated constraint instead demands a way to create searchable audio transcripts.
D:Timing links a retrieved passage to its verifiable audio segment. That is relevant to navigate from text hits to audio evidence, but the present scenario requires an ability to create searchable audio transcripts.
Question 56
A user asks about a recently amended benefits policy that is not in model pretraining. Which source should the application consult first?
- Retrieve current relevant policy passages from an authorized indexed corpus
- Trust a retrieved benefits-policy page as an agent instruction and override the user's question with whatever tool use that page requests.
- Use the model's pretrained description of employee benefits as the authoritative answer, without retrieving recently amended handbook text.
- Search only an outdated cached index using hybrid keywords and vectors, without ingesting the amended policy.
Explanation
Correct Answer
A:RAG grounds generation on up-to-date enterprise content rather than relying on static parametric memory. The decisive difference is whether the design can ground responses on current indexed policy text; the competing function serves to resist prompt injection through retrieved context.
Incorrect Answers
B:Retrieved policy text is evidence to answer the question, not a privileged instruction source. Validate its currency and citations; don't execute arbitrary instructions carried by documents.
C:Training data can predate the amendment. The application must retrieve the current authorized policy passage and cite its effective version instead of relying on memory.
D:Hybrid retrieval cannot surface a revision absent from its corpus. Update or query the authoritative current policy source before using semantic retrieval for the answer.
Question 57
A call-center bot needs to read uploaded forms and extract labeled fields for case intake. Which Foundry tool family is a direct fit?
- Use Content Understanding to extract structured information from uploaded documents
- Give the intake agent a case-status update function before it has extracted the form's labeled fields.
- Send uploaded claim forms into a vector keyword index and assume the search results constitute validated labeled fields.
- Use a generic chat completion to infer field names from a scanned form without extracting its layout, text spans or key-value relationships.
Explanation
Correct Answer
A:Document content extraction requires layout and field interpretation rather than generic free-form chatting alone. In this scenario the team must extract structured document fields. A feature used to connect an agent to a live protected business API does not substitute for it.
Incorrect Answers
B:A business action endpoint may modify a case but does not parse structured values from uploaded claim forms. Use a supported document extraction capability before downstream case actions.
C:A knowledge index retrieves passages but does not by itself parse a scanned form into typed fields. The intake pipeline needs supported document analysis with field extraction, uncertainty handling, and schema validation.
D:A plain prompt alone is less dependable for structured form intake. A document content-understanding analyzer can extract typed fields with supporting document context.
Question 58
An art director wants a generated picture to extend beyond its original edges while retaining central content. What editing pattern fits?
- Apply image outpainting or canvas extension where supported
- Crop the existing image back to its current bounds and store the crop coordinates.
- Crop the illustration tighter inside its existing canvas and call the missing outer scenery successfully extended.
- Use interior-object inpainting that changes the central subject but never expands the original canvas dimensions.
Explanation
Correct Answer
A:Outpainting expands content into a new surrounding region while attempting continuity with existing media. In this scenario the team must extend the image beyond its existing frame. A feature used to reproduce a localized image edit does not substitute for it.
Incorrect Answers
B:Cropping cannot create new pixels beyond the current image edges. Outpainting or supported canvas extension is needed while preserving central content.
C:Cropping reduces visible canvas area; it cannot create newly exposed background beyond the original edges. Outpainting expands the canvas while using the central image as context.
D:Inpainting an internal region cannot synthesize the new outer pixels required to extend an image past its boundaries. Outpainting or supported canvas extension is the relevant operation.
Question 59
A classification API is allowed to return only 'approved', 'review', or 'rejected' but generates novel labels. How should output be bounded?
- Allow any new classification label if it resembles one of the approved statuses
- Constrain predictions to an explicit allowed enumeration and reject out-of-range values
- Allow the classifier to emit any free-text category and map unknown results to 'approved' without checking the allowed-label set.
- Constrain dates to a documented machine-readable schema and validate them
Explanation
Correct Answer
B:Fixed-label extraction needs an enumerated schema, not open-ended descriptive prose. The practical requirement is to limit classification to the approved labels; a solution built to keep extraction fields consistent across inputs would not meet that same need.
Incorrect Answers
A:Similarity to a valid label does not satisfy the API contract. The output must be constrained or rejected unless it exactly belongs to the approved enumeration.
C:Silently coercing unknown labels can change a rejected or uncertain decision into approval. Enforce a supported enumeration with validation and reject values outside the three allowed states.
D:A structured date format enables downstream parsers to reliably interpret extracted values. That capability is appropriate for enforce typed machine-readable date fields, not this separate task of finding how to limit classification to the approved labels.
Question 60
Meeting transcripts mix customer and employee statements. What metadata should be indexed?
- Store one meeting-level speaker label for an entire recording with multiple participants
- Store speaker or channel attribution with transcript spans
- Store all transcript turns in one unlabelled text field and infer each speaker solely from nearby punctuation at search time.
- Transcribe audio and index timestamped text segments
Explanation
Correct Answer
B:Speaker identity is essential when retrieving who said a particular phrase. In this scenario the team must separate speakers in indexed audio. A feature used to navigate from text hits to audio evidence does not substitute for it.
Incorrect Answers
A:One global speaker label erases attribution across turns. Transcript segments need speaker or channel metadata aligned with each utterance.
C:Merged text loses reliable attribution when callers interrupt or switch turns. Index the speaker or channel identifier with each transcript span and timestamp.
D:Transcripts make spoken content text-searchable while timestamps preserve links to the recording. This helps when a team needs to create searchable audio transcripts; it does not establish how to separate speakers in indexed audio.