Microsoft AI-103 Practice Test 15
Question 140
A user asks a natural-language policy question using words not present in the relevant source. What retrieval mechanism helps?
- Require every user query to repeat the source's exact technical vocabulary before permitting a result to match.
- Build a diverse labeled query set covering short and long phrasing
- Apply semantic reranking to top retrieval candidates
- Use semantic retrieval or vector search over appropriately embedded passages
Explanation
Correct Answer
D:Semantic similarity can surface conceptually related text despite vocabulary mismatch. This resolves the specific need to retrieve paraphrased source passages, while a measure that helps score accuracy of top search results addresses a different requirement.
Incorrect Answers
A:A natural-language paraphrase may express the same meaning with different words. Semantic or vector retrieval can bridge that wording gap.
B:Relevance assessment needs representative query diversity before claiming robust semantic quality. That is relevant to evaluate search across different query phrasings, but the present scenario requires an ability to retrieve paraphrased source passages.
C:Reranking can improve the order of already retrieved passages based on query relevance. This helps when a team needs to place precise evidence above generic matches; it does not establish how to retrieve paraphrased source passages.
Question 141
A new response policy produces fewer unsafe outputs but refuses many allowed customer requests. What paired measurements are necessary?
- Report only the number of unsafe outputs blocked and ignore legitimate user requests incorrectly rejected by the new filter.
- Measure the proportion of supported citations while ignoring newly refused legitimate customer requests.
- Calibrate the rubric and compare judgments with human-reviewed anchor examples
- Measure both false-negative safety escapes and false-positive refusals
Explanation
Correct Answer
D:Moderation improvements require balancing actual harms against excessive blocking. Here the relevant capability must let the team evaluate safety overblocking and missed harms. An alternative targeting test generalization beyond tuning cases solves another problem.
Incorrect Answers
A:A lower escape rate is not enough when false refusals rise. Evaluation must measure both harmful content that slips through and allowable content blocked in error.
B:Groundedness can improve without detecting safety overblocking. The release must measure both missed harmful outputs and inappropriate refusals of allowed requests across a representative policy test set.
C:Repeatability and anchored scoring improve confidence in automatic evaluation. That capability is appropriate for control variance in model-based evaluation, not this separate task of finding how to evaluate safety overblocking and missed harms.
Question 142
A regulated claims agent needs current policy answers and durable user-approved preferences but must never expose another user's records. Which THREE design controls address these distinct concerns? Choose THREE.
- Give the claims agent a global action tool that reads all customers' records instead of enforcing policy retrieval and per-user access.
- Use one shared cross-customer conversation history without verifying the current user.
- Enforce document-level authorization before evidence enters agent context
- Connect the agent to a permission-filtered retrieval knowledge source
- Store explicit durable preferences in a governed profile store, excluding secrets
- Use short-lived session state for durable approved preferences while ignoring other users' document ACLs.
Explanation
Correct Answers
C:Retrieval security must use the requester's entitlements; a citation should not reveal restricted text.
D:Knowledge retrieval supplies current citations on demand while avoiding huge prompts and unauthorized corpus exposure.
E:Long-lived user preferences need a controlled source with consent and deletion policies rather than raw unbounded chat logs.
Incorrect Answers
A:A broad action endpoint violates the cross-user privacy condition. Separate current policy retrieval, governed preferences and record authorization without privileged global access.
B:Sharing conversation memory across customers without binding user identity breaches the requirement to isolate authorized records.
F:Session-only memory doesn't meet the durable-preference requirement and cannot protect cross-user records. Use governed consented preference storage plus authorization-filtered retrieval.
Question 143
A disaster-response analyst needs a detailed account of objects, positions and visible damage across three photos. What generation setup fits?
- Describe only the first photo and claim it also represents the damage and positions in the two remaining images.
- Send the image and question to a multimodal visual-understanding model
- Provide all permitted images and request structured detailed observations per image
- Prompt for a concise factual image caption centered on salient visible content
Explanation
Correct Answer
C:Each permitted image should be inspected independently with structured observations of visible damage, location and uncertainty. Aggregate only after preserving the image-to-finding association; a generic single caption cannot support a defensible multi-image damage report.
Incorrect Answers
A:Three distinct photographs can show different evidence. Multimodal analysis must inspect each image, preserve attribution and compare the observed features rather than generalizing one frame.
B:The model must inspect the visual scene rather than answer from the user's words alone. The function serves to answer an image-grounded visual count, whereas the actual requirement is to describe several images with source separation.
D:A short factual caption should describe what is visible without unrelated speculation. Here the needed outcome is to describe several images with source separation; an option designed to produce compact visually grounded captions would not directly provide it.
Question 144
An internal moderation model blocks harmless chemistry questions because they include flagged words. What analysis improvement helps?
- Evaluate contextual safety classifications rather than using keywords alone
- Continue blocking chemistry messages using an isolated list of hazardous keywords
- Block all messages containing a chemistry keyword regardless of whether the surrounding explanation is educational and harmless.
- Detect and redact protected personal identifiers before sharing the output
Explanation
Correct Answer
A:Unsafe intent depends on context; simple word matches produce overblocking. This directly supports the need to reduce false-positive text moderation, unlike a mechanism primarily intended to escalate threatening content for human review.
Incorrect Answers
B:Keyword overlap alone does not establish disallowed intent; legitimate educational questions can mention hazardous substances. Context-sensitive safety evaluation reduces false positives.
C:Keyword-only safety screening cannot distinguish context. Evaluate intent and surrounding text to reduce false positive refusals while retaining safety protections.
D:Sensitive-data recognition and masking reduce privacy leakage from otherwise useful summaries. Its primary purpose is to protect sensitive identifiers in text outputs; the stated constraint instead demands a way to reduce false-positive text moderation.
Question 145
A search index returns semantically related passages but ranks a broad introduction over a precise answer. What additional stage may help?
- Increase the vector result count without reranking generic passages against the actual question
- Return candidates strictly in their raw vector similarity order even when an exact answer passage is ranked below broad introductory content.
- Apply semantic reranking to top retrieval candidates
- Use semantic retrieval or vector search over appropriately embedded passages
Explanation
Correct Answer
C:Reranking can improve the order of already retrieved passages based on query relevance. The practical requirement is to place precise evidence above generic matches; a solution built to evaluate search across different query phrasings would not meet that same need.
Incorrect Answers
A:Retrieving more loosely related introductions may worsen evidence focus. Ranking must distinguish specific passages relevant to the user's question from broadly similar text.
B:Initial similarity is a coarse retrieval signal. Semantic reranking can compare the candidate passages to the full question and lift the answer-bearing evidence.
D:Semantic similarity can surface conceptually related text despite vocabulary mismatch. That capability is appropriate for retrieve paraphrased source passages, not this separate task of finding how to place precise evidence above generic matches.
Question 146
An AI service records a user's private access token in a model-visible tool result. How should tool output be transformed?
- Pass the complete tool response through and ask the LLM to ignore token fields.
- Base64-encode the access token before adding it to the model's message.
- Store the raw authorization header as an additional tool argument for transparency.
- Redact credentials from tool responses and expose only task-relevant structured fields to the model.
Explanation
Correct Answer
D:Secrets passed into the prompt can be repeated, logged or leaked; the controller should minimize tool-result disclosure.
Incorrect Answers
A:An instruction to ignore a secret leaves the credential inside model context and traces.
B:Encoding does not remove the secret from model-visible data or prevent its recovery.
C:Exposing the authorization header to downstream tools increases credential leakage risk.
Question 147
A product team needs to isolate development AI assets from production access and billing. Which infrastructure boundary should be planned?
- Store development and production projects in one shared privileged resource context while separating only their logging dashboards.
- Apply one development tagging policy but continue sharing the same Foundry project identity and production resource group.
- Use separate appropriately governed projects and resource scopes for environments
- Deploy a second stateless controller in the same development resource group and grant it production administration.
Explanation
Correct Answer
C:Environment separation limits accidental cross-environment changes and supports distinct access and cost controls. The decisive difference is whether the design can isolate development and production resources; the competing function serves to retain operational telemetry independent of app instances.
Incorrect Answers
A:Observability separation does not isolate access rights, cost or destructive changes. Use governed project and resource boundaries for development versus production.
B:Tag compliance cannot establish independent authorization or billing boundaries when environments share access-bearing resources. Separate governed projects and scopes for development and production.
D:Replication within one shared environment does not isolate development access or billing. Use governed production and development project/resource boundaries, distinct identities and environment-specific policies.
Question 148
A model claims a blurry label says a serial number that cannot actually be read. What should the response policy require?
- Force a precise serial number by majority-voting guesses from the same blurry pixels, without recording uncertainty.
- Transcribe the unreadable serial as a confident exact value because the image is accepted by the vision model.
- Verify chart-reading accuracy including axes, units and legends
- Express uncertainty when the image lacks sufficient resolution to support recognition
Explanation
Correct Answer
D:Visual evidence limits must govern claims; hallucinating obscured text can cause operational errors. This directly supports the need to avoid inventing unreadable visual details, unlike a mechanism primarily intended to describe several images with source separation.
Incorrect Answers
A:Repeated guesses do not create missing image resolution. When characters are unreadable, the response should mark the value uncertain and request a clearer image rather than fabricate an exact identifier.
B:An API accepting the image does not make illegible pixels legible. The response must disclose uncertainty rather than inventing text that cannot be resolved.
C:Visual numeric interpretation requires more than recognizing colored bars or lines. Here the needed outcome is to avoid inventing unreadable visual details; an option designed to interpret chart values with labels and units would not directly provide it.
Question 149
A translator receives invoices with product codes such as XG-440 and must not translate them. What should be configured?
- Translate every character including SKU codes into a target-language approximation.
- Remove all product identifiers before generating the translation.
- Preserve protected named entities and placeholders through translation with post-translation checks.
- Force sentence-level fluency without comparing identifiers after translation.
Explanation
Correct Answer
C:Code identifiers are not prose. Glossaries and protected spans prevent unintended substitution or localization.
Incorrect Answers
A:Altering exact product identifiers prevents the translated invoice from matching the source inventory records.
B:Deleting identifiers loses business-critical references and does not preserve them accurately.
D:Fluent output can still corrupt SKU strings unless protected tokens are checked.
Question 150
A search index returns archived policies even after records are replaced with a new signed version. What index update strategy is required?
- Retain all versions with no effective-date or active-state filtering.
- Rebuild only the vector embeddings and ignore the source version mapping.
- Change search-result ranking by document length rather than state.
- Use stable document IDs with change detection and retire superseded chunks from search.
Explanation
Correct Answer
D:Incremental ingestion must update and remove old versions, not leave duplicates indefinitely searchable.
Incorrect Answers
A:Obsolete records can remain retrievable and contaminate answers even after a new policy is signed.
B:Recomputing embedding vectors does not remove outdated policy versions or change which source revision is effective. The indexer must reconcile original document IDs and mark superseded chunks unavailable for current retrieval.
C:Length cannot distinguish an approved current policy from an archived revision.