Microsoft AI-103 Practice Test 4
Question 30
A search index has empty content for scans of signed paper forms. Which enrichment should run first?
- Preserve OCR table-cell structure and bounding regions during extraction
- Run semantic reranking on the empty searchable text field of the scanned pages without first recognizing the characters.
- Use layout-aware reading order and column segmentation
- Apply optical character recognition to scanned image pages
Explanation
Correct Answer
D:OCR converts visual text into searchable characters when no text layer is available. This resolves the specific need to extract text from scanned pages, while a measure that helps retain structured tabular data from images addresses a different requirement.
Incorrect Answers
A:Table-aware extraction maintains which values belong to which headings and rows. Its primary purpose is to retain structured tabular data from images; the stated constraint instead demands a way to extract text from scanned pages.
B:A reranker cannot recover words absent from the index. Scanned paper forms require OCR to produce text before passage ranking can work.
C:Page layout structure is necessary for sensible text order in multi-column documents. This helps when a team needs to preserve reading order on complex pages; it does not establish how to extract text from scanned pages.
Question 31
A model answer is cut off because the generated response hits the output-token limit. What request setting must be reconsidered?
- Switch to JSON-only output while leaving the generated-token budget below the size of the required response.
- Change which deployment receives the request but keep the response-token limit too small.
- Change the transport from HTTPS to a streaming response while leaving the maximum generated-token allowance below the required answer length.
- Adjust the response output-token budget while controlling required field size
Explanation
Correct Answer
D:Truncation can arise from an insufficient maximum generation allowance; increasing it blindly can raise costs. This resolves the specific need to allow enough output tokens for the required answer, while a measure that helps enforce a structured response shape addresses a different requirement.
Incorrect Answers
A:A JSON constraint may enforce fields but will not prevent truncation when the permitted output token budget is too small. Inspect finish reasons and use a suitable supported generation allowance.
B:Changing deployment names does not address a response truncated by its generation token budget. Inspect and adjust the supported output allowance.
C:Streaming changes delivery mechanics but does not enlarge a model's permitted completion budget. Check the finish reason and adjust supported output limits appropriately.
Question 32
A mobile assistant must produce fluent natural-language responses but may retrieve current product prices from a database. How should model capabilities be divided?
- Use a text-only model to invent current prices from pretraining and avoid querying the live pricing data source.
- Use a language model for the response and a trusted retrieval tool for current prices
- Select a multimodal model that accepts images together with text
- Test a smaller model against the fixed-label accuracy target before selecting a large model
Explanation
Correct Answer
B:Generation handles natural language, while the product database remains the authoritative source for facts that change frequently. The practical requirement is to separate language generation from live authoritative data; a solution built to recognize object content in images would not meet that same need.
Incorrect Answers
A:Language generation cannot guarantee current catalog prices. A trusted retrieval or API call must supply authoritative price data to the response workflow.
C:A text-only model cannot directly inspect pixels; a multimodal input model can jointly use the submitted image and the accompanying question. The function serves to interpret images and text in one request, whereas the actual requirement is to separate language generation from live authoritative data.
D:Structured narrow tasks may not require a general large model; accuracy, latency and measured spend should determine the smallest adequate choice. That capability is appropriate for pick sufficient capability for a bounded output taxonomy, not this separate task of finding how to separate language generation from live authoritative data.
Question 33
An image-generation service is expensive and marketing repeatedly creates almost identical assets. What production optimization helps?
- Generate the same approved product image on every request with a stricter image quality setting, without reusing any stored approved asset.
- Use a supported reference-image-conditioned generation workflow
- Select an available text-to-video generation capability and suitable output constraints
- Cache approved outputs and reuse deterministic asset identifiers when licensing permits
Explanation
Correct Answer
D:Reusing approved visual outputs avoids unnecessary repeated generation and improves brand consistency. The practical requirement is to avoid redundant media generation requests; a solution built to control supported video presentation parameters would not meet that same need.
Incorrect Answers
A:Increasing output quality can make repeated generation more expensive; it does not eliminate identical work. Cache approved outputs and reuse governed asset identifiers when allowed.
B:Reference media can condition composition or appearance in ways that text-only instructions may not capture reliably. This helps when a team needs to condition image generation on reference visual material; it does not establish how to avoid redundant media generation requests.
C:Video generation produces a temporal sequence; a still-image API will not produce moving output on its own. The function serves to create moving video clips from prompts, whereas the actual requirement is to avoid redundant media generation requests.
Question 34
A call log summary must preserve what the customer requested and what the agent promised. What prompt requirement helps?
- Collapse the customer's request and the agent's promise into one untyped summary field
- Require separate factual fields for customer requests and agent commitments
- Publish the call summary with personal identifiers intact while omitting the customer's request and the agent's commitments.
- Generate a source-grounded summary with cited contract sections
Explanation
Correct Answer
B:Structured summary sections keep the two parties' actions distinguishable. This directly supports the need to preserve speaker commitments in conversation summaries, unlike a mechanism primarily intended to include important evidence located outside opening sections.
Incorrect Answers
A:A single mixed field obscures who requested what and which commitments were actually made. The output needs distinct, attributable facts to support follow-up and audit.
C:Merely retaining sensitive data does not capture the operational promises or task facts. Use separate grounded fields for what was requested and what the agent agreed to do, applying privacy protection separately.
D:Grounding with passage references helps reviewers verify that reported obligations occur in the document. Its primary purpose is to summarize a long contract with traceable evidence; the stated constraint instead demands a way to preserve speaker commitments in conversation summaries.
Question 35
An OCR pipeline merges two newspaper columns into one scrambled sentence. What additional extraction is needed?
- Use layout-aware reading order and column segmentation
- Sort detected words only by vertical page coordinate across both newspaper columns
- Treat a two-column newspaper as one continuous left-to-right text line across both columns.
- Validate extracted identifiers and route uncertain OCR outputs for review
Explanation
Correct Answer
A:Page layout structure is necessary for sensible text order in multi-column documents. The practical requirement is to preserve reading order on complex pages; a solution built to extract text from scanned pages would not meet that same need.
Incorrect Answers
B:A page-global top-to-bottom sort interleaves two independent columns. Layout-aware column detection and reading-order recovery are required before reconstructing sentences.
C:Cross-column reading merges unrelated sentences. OCR must respect page layout and reading order to reconstruct each column independently.
D:Uncertain recognition should not silently become authoritative structured data. That capability is appropriate for avoid accepting incorrect low-confidence OCR, not this separate task of finding how to preserve reading order on complex pages.
Question 36
A search assistant suddenly sends hundreds of parallel generation calls and gets throttled. What client feature should it add?
- Use bounded concurrency and retry-after-aware backoff for model requests
- Send the same unbounded burst directly to the deployment without concurrency control or retry-after-aware backoff.
- Increase the application's concurrency without respecting rate-limit responses or retry-after guidance.
- Adjust the response output-token budget while controlling required field size
Explanation
Correct Answer
A:Client-side admission and backoff prevent synchronized floods during provider throttling. The practical requirement is to manage model request concurrency and throttling; a solution built to invoke the configured model endpoint would not meet that same need.
Incorrect Answers
B:A valid model endpoint alone cannot prevent throttling from excessive parallel requests. Bound concurrency and follow rate-limit feedback.
C:More unbounded concurrent calls aggravate throttling. Cap in-flight requests and apply provider-aware backoff with appropriate retry budgets.
D:Truncation can arise from an insufficient maximum generation allowance; increasing it blindly can raise costs. That capability is appropriate for allow enough output tokens for the required answer, not this separate task of finding how to manage model request concurrency and throttling.
Question 37
A photo tagging tool must identify several objects in a picture, not write a sales narrative. What should the architecture prioritize?
- Select a visual-understanding model assessed on object-level recognition
- Validate the system only against a text medical-abbreviation set and assume that proves object localization in photographs.
- Generate a sales caption from a product catalog without detecting or locating the several visible objects in the customer photo.
- Choose a general marketing-caption model because its prose reads well, without testing whether it detects the individual visible objects.
Explanation
Correct Answer
A:Object discovery is a perception problem; evaluate image understanding outputs rather than relying on prose-generation fluency alone. Here the relevant capability must let the team recognize object content in images. An alternative targeting evaluate model performance on representative domain input solves another problem.
Incorrect Answers
B:Text classification evaluation does not test recognizing multiple visible objects in a photo. Use a visual understanding model and representative annotated images for the actual tagging workload.
C:Catalog retrieval and natural-language generation do not by themselves provide object-level recognition. Test a visual understanding model with representative annotated images and required object tags.
D:A fluent caption can omit or invent objects. Object-level image tagging should be evaluated on representative annotated photos using a visual recognition capability.
Question 38
A preview generated image lacks traceability to its prompt and source references. What provenance should be recorded?
- Discard the source and prompt identifiers whenever a preview asset has a safety warning.
- Keep only the latest output cache key and delete all prompt versions, reference identifiers and generating request records.
- Archive the final image URL alone while discarding the generation prompt version, input reference IDs and producing request identifier.
- Store generation request IDs, prompt versions, input references and output hashes
Explanation
Correct Answer
D:Provenance supports reproducibility, moderation review and asset management. Here the relevant capability must let the team trace synthetic imagery back to generating inputs. An alternative targeting handle disallowed generated-image requests safely solves another problem.
Incorrect Answers
A:Discarding identifiers prevents later audit of the generated image. Preserve permitted version, source and prompt metadata while respecting safety refusals.
B:A cache key identifies stored content but cannot reconstruct the specific generation inputs and version history. Provenance requires linkage among request IDs, input references, prompt versions and output hashes.
C:A bare asset URL cannot explain which prompt or reference media created an image. Provenance requires the generating request and source identity alongside its output digest.
Question 39
A classifier marks a sarcastic complaint as positive because it contains the word 'wonderful.' What should be evaluated?
- Classify sentiment by counting positive words such as wonderful.
- Lower the model's generation temperature without adding sarcasm cases to evaluation.
- Use context-sensitive sentiment evaluation and sarcasm-specific held-out examples.
- Treat every exclamation mark as evidence of satisfaction.
Explanation
Correct Answer
C:Keyword polarity alone can invert customer intent when sarcasm changes the meaning of positive words.
Incorrect Answers
A:Lexical polarity alone fails when sarcasm reverses the apparent meaning of a positive token.
B:Lower sampling variability does not teach a classifier to interpret sarcasm in context.
D:Punctuation is ambiguous and cannot distinguish an ironic complaint from genuine praise.
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.