Microsoft AI-103 Practice Test 23
Question 220
A hybrid result set is dominated by near-duplicate snippets from one manual. What post-retrieval processing helps?
- Raise top-k while allowing the same manual to occupy every result slot
- Use compatible embedding representations for indexed documents and search queries
- Deduplicate or diversify ranked results while retaining strongest evidence
- Align embedding output dimension with the index vector field definition
Explanation
Correct Answer
C:Source diversity and duplicate filtering improve usable coverage for grounded generation. The practical requirement is to avoid redundant evidence in hybrid retrieval; a solution built to match exact identifiers plus related concepts would not meet that same need.
Incorrect Answers
A:Increasing result count does not address redundancy from one document. Diversification or grouping by source is needed to improve coverage of different evidence.
B:Vector distances are meaningful only when query and document vectors inhabit compatible spaces. This helps when a team needs to keep document and query embeddings compatible; it does not establish how to avoid redundant evidence in hybrid retrieval.
D:Vector search requires vectors whose shape agrees with the configured schema. That capability is appropriate for prevent vector field shape mismatches, not this separate task of finding how to avoid redundant evidence in hybrid retrieval.
Question 221
An agent receives a user request without essential account context and proceeds with a guess. What behavior should be designed?
- Tag each conversation and tool response with stable isolated session and task identifiers
- Maintain a short ordered conversation and approval state tied to the task
- Minimize conversation memory and redact secrets before persistence or replay
- Ask for missing required input before invoking account-changing tools
Explanation
Correct Answer
D:Clarifying mandatory parameters prevents incorrect side effects based on invented assumptions. In this scenario the team must collect missing mandatory task context. A feature used to isolate concurrent conversation state does not substitute for it.
Incorrect Answers
A:Correlation IDs prevent results and memory from leaking across concurrent requests. Here the needed outcome is to collect missing mandatory task context; an option designed to isolate concurrent conversation state would not directly provide it.
B:Relevant conversation state keeps authorization and progress coherent across tool cycles. Its primary purpose is to retain task-scoped user decisions during turns; the stated constraint instead demands a way to collect missing mandatory task context.
C:Retaining only necessary state reduces exposure of sensitive credentials in future prompts. That is relevant to keep unnecessary secrets out of agent memory, but the present scenario requires an ability to collect missing mandatory task context.
Question 222
An engineer wants to track which prompt-template commit produced an incorrect response. What pipeline metadata should be logged?
- Run a representative evaluation suite with defined failure thresholds before promotion
- Record only an infrastructure-as-code template revision while omitting the prompt-template commit, deployment ID and request trace.
- Attach prompt, model and application version identifiers to release and trace records
- Require a recorded protected-environment approval gate for safety changes
Explanation
Correct Answer
C:Version provenance makes regression triage and controlled rollback possible. In this scenario the team must trace behavior to the precise released configuration. A feature used to test model behavior before promotion does not substitute for it.
Incorrect Answers
A:A measurable evaluation gate prevents known regressions from reaching users. Here the needed outcome is to trace behavior to the precise released configuration; an option designed to test model behavior before promotion would not directly provide it.
B:Infrastructure revisions cannot identify which prompt text actually produced a response. Log the deployed prompt-template version or commit alongside model deployment and request identifiers for reproducible attribution.
D:Release approvals must be checked by the pipeline, not just requested in a prompt. This helps when a team needs to enforce auditable approval of safety configuration; it does not establish how to trace behavior to the precise released configuration.
Question 223
An alt text generator writes 'image of' before every output. What style guidance should be added?
- Describe meaningful visual content directly without redundant introductory phrases
- Write context-specific descriptions of each step rather than repeating one generic caption
- Include relevant categories, trends, units and critical numeric comparisons
- Prefix every image description with the words image of even when context already conveys that an image is present.
Explanation
Correct Answer
A:Screen readers already convey the image role, so descriptive content should carry the information. The decisive difference is whether the design can write concise direct alt descriptions; the competing function serves to distinguish meaningful steps across several images.
Incorrect Answers
B:Per-image alternatives let nonvisual users follow the sequence. Here the needed outcome is to write concise direct alt descriptions; an option designed to distinguish meaningful steps across several images would not directly provide it.
C:A graph alternative should communicate the information conveyed visually, not merely its colors. Its primary purpose is to carry chart quantitative meaning into text; the stated constraint instead demands a way to write concise direct alt descriptions.
D:Repeated image-of prefixes often add noise without describing useful content. Alt text should communicate function or meaning directly and avoid unnecessary labels.
Question 224
A compliance department needs to detect obligations and deadlines in company policies. What output design fits?
- Define obligation, responsible party and deadline fields for domain extraction
- Add modality-sensitive domain examples distinguishing permissions from obligations
- Represent unsupported fields as null or unavailable rather than fabricating values
- Use representative domain reports with explicitly labeled exception examples
Explanation
Correct Answer
A:A domain-specific schema enables structured auditing rather than generic sentiment summaries. The decisive difference is whether the design can extract actionable compliance obligations; the competing function serves to interpret domain-specific modal language.
Incorrect Answers
B:The difference between optional and mandatory terms is material to legal task accuracy. Here the needed outcome is to extract actionable compliance obligations; an option designed to interpret domain-specific modal language would not directly provide it.
C:Extractive business applications must not invent values absent from the source. Its primary purpose is to represent missing evidence explicitly in domain extraction; the stated constraint instead demands a way to extract actionable compliance obligations.
D:Domain evidence and exception cases let teams test whether critical caveats survive summarization. That is relevant to evaluate finance-specific exception handling, but the present scenario requires an ability to extract actionable compliance obligations.
Question 225
A search index must include detected entities from uploaded contracts in addition to raw text. Which pipeline operation fits?
- Configure entity extraction as an enrichment step before indexing
- Index the contract's title and filename only, then assume legal entities will be available as structured searchable fields.
- Add document-field or key-value extraction using supported content analyzers
- Verify output field mappings and target index schema compatibility
Explanation
Correct Answer
A:Index-time enrichment adds structured searchable fields derived from source content. Here the relevant capability must let the team enrich index records with extracted entities. An alternative targeting extend indexing with custom classification solves another problem.
Incorrect Answers
B:File metadata is not equivalent to extracting entities from contract content. An enrichment operation must detect the relevant spans and persist them in mapped index fields.
C:Field-aware extraction makes important structured document attributes accessible to filters or agents. That capability is appropriate for index key-value form fields, not this separate task of finding how to enrich index records with extracted entities.
D:Enrichment results must be explicitly mapped to correctly typed index fields. Here the needed outcome is to enrich index records with extracted entities; an option designed to map enrichment outputs into indexed fields would not directly provide it.
Question 226
A billing agent needs a memory of customer preferences across sessions but cannot retain payment-card data. Which design is appropriate?
- Store complete conversation transcripts including credit-card fields indefinitely.
- Save only a global preference profile shared among all customers.
- Persist only approved preference fields with retention policy and explicit exclusion of sensitive payment details.
- Require users to reenter and log card numbers every new session.
Explanation
Correct Answer
C:Long-term memory must follow consent and data-minimization rules; unrestricted transcript persistence is unsafe.
Incorrect Answers
A:Unbounded raw storage violates the explicit exclusion of payment-card data and retention minimization.
B:A single shared preference profile can mix customer histories and expose private information to other users. Store only permitted preferences under a customer-specific identity and exclude payment credentials from persistent memory.
D:Repeatedly collecting and retaining card details contradicts the stated privacy requirement.
Question 227
CI tests pass, but deployed agents fail because a required external tool endpoint is unreachable. What test should be added?
- Run a representative evaluation suite with defined failure thresholds before promotion
- Attach prompt, model and application version identifiers to release and trace records
- Declare supported AI infrastructure and deployment settings as version-controlled automation
- Run post-deployment connectivity and permission smoke tests for critical tools
Explanation
Correct Answer
D:Build tests cannot validate environment-specific tool networking and authorization. This resolves the specific need to validate actual deployed tool dependencies, while a measure that helps test model behavior before promotion addresses a different requirement.
Incorrect Answers
A:A measurable evaluation gate prevents known regressions from reaching users. Its primary purpose is to test model behavior before promotion; the stated constraint instead demands a way to validate actual deployed tool dependencies.
B:Version provenance makes regression triage and controlled rollback possible. That is relevant to trace behavior to the precise released configuration, but the present scenario requires an ability to validate actual deployed tool dependencies.
C:Infrastructure-as-code reduces drift and makes resource changes reviewable and reproducible. This helps when a team needs to reproduce deployment infrastructure across environments; it does not establish how to validate actual deployed tool dependencies.
Question 228
A chart's key finding depends on units and axis direction. What information must accessible text preserve?
- Describe meaningful visual content directly without redundant introductory phrases
- Describe the chart as a colorful graphic while omitting axis units and critical numeric comparisons.
- Review generated alternatives against the final published image versions
- Include relevant categories, trends, units and critical numeric comparisons
Explanation
Correct Answer
D:A graph alternative should communicate the information conveyed visually, not merely its colors. In this scenario the team must carry chart quantitative meaning into text. A feature used to write concise direct alt descriptions does not substitute for it.
Incorrect Answers
A:Screen readers already convey the image role, so descriptive content should carry the information. Here the needed outcome is to carry chart quantitative meaning into text; an option designed to write concise direct alt descriptions would not directly provide it.
B:Presentation-only language does not convey the chart's data. Accessible text must preserve the trend, units, direction and important comparisons.
C:Text alternatives must correspond to what users actually see after edits or revisions. That is relevant to keep accessibility descriptions aligned with final media, but the present scenario requires an ability to carry chart quantitative meaning into text.
Question 229
A financial report summarizer must preserve exceptions that impact audit findings. What training/evaluation corpus matters?
- Drop exception paragraphs so the generated compliance summary is shorter
- Add modality-sensitive domain examples distinguishing permissions from obligations
- Use representative domain reports with explicitly labeled exception examples
- Represent unsupported fields as null or unavailable rather than fabricating values
Explanation
Correct Answer
C:Domain evidence and exception cases let teams test whether critical caveats survive summarization. In this scenario the team must evaluate finance-specific exception handling. A feature used to extract actionable compliance obligations does not substitute for it.
Incorrect Answers
A:An exception can change the interpretation of a policy obligation. Removing it for brevity loses material audit facts and can misrepresent applicability.
B:The difference between optional and mandatory terms is material to legal task accuracy. Its primary purpose is to interpret domain-specific modal language; the stated constraint instead demands a way to evaluate finance-specific exception handling.
D:Extractive business applications must not invent values absent from the source. This helps when a team needs to represent missing evidence explicitly in domain extraction; it does not establish how to evaluate finance-specific exception handling.
Question 230
An indexer extracts raw page text but doesn't preserve key-value pairs from structured forms. What enrichment should be considered?
- Index only plain text and discard extracted field names and their associated values
- Configure entity extraction as an enrichment step before indexing
- Invoke a custom enrichment skill with a defined request and response contract
- Add document-field or key-value extraction using supported content analyzers
Explanation
Correct Answer
D:Field-aware extraction makes important structured document attributes accessible to filters or agents. This directly supports the need to index key-value form fields, unlike a mechanism primarily intended to map enrichment outputs into indexed fields.
Incorrect Answers
A:Losing key-value structure prevents field-aware lookup and reliable verification. Rich enrichment should preserve labels and values as explicit index fields.
B:Index-time enrichment adds structured searchable fields derived from source content. That capability is appropriate for enrich index records with extracted entities, not this separate task of finding how to index key-value form fields.
C:Custom skills insert application-owned processing into the ingestion workflow. Here the needed outcome is to index key-value form fields; an option designed to extend indexing with custom classification would not directly provide it.