cert
cert-1
cert-2

Pass Microsoft AI-103 Exam in First Attempt Guaranteed!

Get 100% Latest Exam Questions, Accurate & Verified Answers to Pass the Actual Exam!
30 Days Free Updates, Instant Download!

cert-5
cert-6
AI-103 Exam - Verified By Experts
AI-103 Premium File

AI-103 Premium File

$89.99
$98.99
  • Premium File 187 Questions & Answers. Last Update: Nov 15, 2026

Whats Included:

  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates
 
$98.99
$89.99
accept 397 downloads in the last 7 days
block-screenshots
Microsoft AI-103 Practice Test 3

Question 20

Large manuals are stored as single index documents and retrieval supplies irrelevant chapters. How should they be partitioned?

  1. Index each complete manual as one vector, then request fewer results without splitting chapters
  2. Split documents into semantically coherent chunks with stable source offsets
  3. Keep each 200-page manual in one giant index entry but assign a stable document ID to it.
  4. Track source deletions and remove or mark deleted index entries

Explanation

Correct Answer

B:

Well-designed chunks improve retrieval specificity while retaining citation lineage. The decisive difference is whether the design can index targeted passages rather than whole manuals; the competing function serves to prepare document text and identity for retrieval.

Incorrect Answers

A:

Changing the top-k count does not create chapter-level retrieval units; irrelevant passages remain embedded in an oversized document. Chunking should respect section boundaries and preserve provenance.

C:

Stable identifiers help updates, but whole-manual chunks still hide the relevant passage among unrelated chapters. Segment by coherent sections with offsets and provenance.

D:

Ingestion must reconcile removed source content so responses do not cite obsolete files. That is relevant to synchronize indexed deletions, but the present scenario requires an ability to index targeted passages rather than whole manuals.

Question 21

Users need predictable machine-readable invoice fields. Which model response constraint should the Python client request?

  1. Apply a supported JSON schema or structured-output response contract
  2. Invoke the model successfully but request unrestricted prose output and ask the invoice importer to guess field types from sentences.
  3. Capture usage counts with the deployment name for per-model cost attribution
  4. Submit a supported multimodal message containing image input and accompanying text

Explanation

Correct Answer

A:

Structured output constraints help downstream parsers consume known required properties and types. In this scenario the team must enforce a structured response shape. A feature used to select the correct model serving API does not substitute for it.

Incorrect Answers

B:

Endpoint connectivity alone does not make an invoice response parseable. A structured-output or schema contract is needed to constrain machine-readable fields and types.

C:

Usage telemetry tied to deployment identifiers supports spend and optimization decisions. That is relevant to account for input and output model tokens, but the present scenario requires an ability to enforce a structured response shape.

D:

The message payload must include the actual visual content in a format accepted by the multimodal API. This helps when a team needs to supply image evidence in model input; it does not establish how to enforce a structured response shape.

Question 22

An internal service only extracts one of six approved issue codes from clean short text. What cost-conscious selection method should be used?

  1. Pick the largest available general model because no cheaper model can ever produce a six-label taxonomy.
  2. Choose the largest available model based solely on a vendor benchmark, without measuring six-label classification accuracy or runtime cost.
  3. Select a visual-understanding model assessed on object-level recognition
  4. Test a smaller model against the fixed-label accuracy target before selecting a large model

Explanation

Correct Answer

D:

Structured narrow tasks may not require a general large model; accuracy, latency and measured spend should determine the smallest adequate choice. In this scenario the team must pick sufficient capability for a bounded output taxonomy. A feature used to handle difficult multi-document reasoning with measured quality does not substitute for it.

Incorrect Answers

A:

A six-label task may be served by a much smaller model. Compare candidates against the actual accuracy requirement, latency and cost rather than assuming scale is necessary.

B:

An unrelated leaderboard cannot show whether a small model meets the actual six-code classification target. Compare candidate models against labeled issue-code examples, cost and response-time requirements.

C:

Object discovery is a perception problem; evaluate image understanding outputs rather than relying on prose-generation fluency alone. That is relevant to recognize object content in images, but the present scenario requires an ability to pick sufficient capability for a bounded output taxonomy.

Question 23

An application sends a prompt that violates configured image-generation policy and the request is denied. How should the client behave?

  1. Automatically resubmit the unchanged disallowed prompt to a second generator after receiving a policy denial.
  2. Return a previously cached copy of the prohibited image after a new generation request is refused by content policy.
  3. Save the blocked request ID and immediately retry the same prohibited prompt without changing policy conditions.
  4. Honor the safety rejection and offer a compliant prompt revision path

Explanation

Correct Answer

D:

Client retries must not bypass provider or enterprise content safety restrictions. In this scenario the team must handle disallowed generated-image requests safely. A feature used to create a synthetic image from prose instructions does not substitute for it.

Incorrect Answers

A:

A blocked request must not be routed around safety controls. The client should preserve the refusal and offer a compliant prompt revision rather than bypassing enforcement.

B:

Reusing a stored asset to evade a refusal undermines the same safety restrictions. The client should honor the rejection and offer a compliant revision path rather than serving disallowed content.

C:

Recording a refusal is useful operationally, but retrying a disallowed request unchanged cannot satisfy safety restrictions. Handle the denial gracefully and request a compliant revision where appropriate.

Question 24

A product taxonomy model often maps 'Galaxy' to a smartphone though users mean astronomy. What input improvement helps?

  1. Resolve every occurrence of Galaxy to the dominant product-catalog sense without local context
  2. Recognize 'Galaxy' as an entity and always assign it to the smartphone category without inspecting nearby words.
  3. Extract named entities and domain-defined identifier fields
  4. Supply local sentence context and domain-specific disambiguation criteria

Explanation

Correct Answer

D:

Ambiguous terms require surrounding evidence and taxonomy rules to resolve the intended sense. The practical requirement is to disambiguate entity meanings using context; a solution built to discover recurring themes without fixed labels would not meet that same need.

Incorrect Answers

A:

A globally frequent product interpretation can be wrong in an astronomy question. Disambiguation must consider surrounding words and the document domain, not just catalog frequency.

B:

The same surface word can refer to consumer electronics or astronomy. Examine local sentence context and domain definitions before assigning the intended meaning.

C:

Entity extraction identifies spans that represent people, organizations and specific domain values. The function serves to identify entity spans and policy identifiers, whereas the actual requirement is to disambiguate entity meanings using context.

Question 25

An indexer ingests the same updated file many times under new identifiers. What identity policy fixes duplication?

  1. Use stable source document IDs and deterministic update semantics
  2. Assign a fresh random document key on every indexer run and suppress only exact text matches
  3. Create a new index record ID on every ingestion and ignore whether the source file has already been indexed.
  4. Extract and normalize document text with source identifiers before indexing

Explanation

Correct Answer

A:

Stable identities let repeated ingestion update an existing indexed record rather than create duplicates. In this scenario the team must deduplicate recurrent document ingestion. A feature used to index targeted passages rather than whole manuals does not substitute for it.

Incorrect Answers

B:

New random keys cause successive versions of one source to appear as unrelated documents. Stable source identity is needed for deterministic update, replacement and deletion.

C:

This behavior causes the reported duplication. Use a stable source identity and deterministic upsert behavior so reingestion replaces the prior entry.

D:

Normalized text and provenance enable later retrieval and citation of relevant policy passages. This helps when a team needs to prepare document text and identity for retrieval; it does not establish how to deduplicate recurrent document ingestion.

Question 26

A model completion includes a plausible answer but no source-backed evidence. What should the RAG controller do before presenting it as confirmed?

  1. Accept the answer because the model's confidence wording sounds certain.
  2. Cite the first retrieved document, even when its contents do not match the claim.
  3. Increase the retrieval result count and publish without examining source-to-claim alignment.
  4. Evaluate claim-level grounding against retrieved authorized passages and withhold unsupported assertions.

Explanation

Correct Answer

D:

A retrieval hit by itself does not prove the generated statement. Evidence should support the actual content of each material claim.

Incorrect Answers

A:

Confident prose cannot substitute for verifying that retrieved evidence entails each factual assertion.

B:

Attaching an irrelevant citation creates false support instead of verifying the claim against the passage.

C:

More retrieved documents do not guarantee that the particular generated assertion is grounded.

Question 27

A team compares AI models solely by vendor leaderboard scores although its own workload contains medical abbreviations. What should they do before deployment?

  1. Select an offline compact model based solely on its size and skip evaluation on the organization's medical abbreviations.
  2. Choose a vision model solely because it leads an unrelated general image benchmark, without testing clinical abbreviations.
  3. Build an application-specific evaluation set containing the medical abbreviations
  4. Use a language model for the response and a trusted retrieval tool for current prices

Explanation

Correct Answer

C:

General benchmark rankings do not demonstrate performance on the organization's domain language and error costs. This resolves the specific need to evaluate model performance on representative domain input, while a measure that helps run short classifications with limited hardware and no network addresses a different requirement.

Incorrect Answers

A:

Running locally might improve latency or privacy, but model size alone cannot establish domain-language accuracy. Build a labeled evaluation set containing actual medical abbreviations and measure errors before choosing a deployment model.

B:

Image benchmark rankings do not establish performance on the application's abbreviated medical text. Use representative domain-specific evaluation data and accuracy criteria.

D:

Generation handles natural language, while the product database remains the authoritative source for facts that change frequently. The function serves to separate language generation from live authoritative data, whereas the actual requirement is to evaluate model performance on representative domain input.

Question 28

A video generator produces the required scene but at the wrong duration and aspect ratio. What should the application set?

  1. Record output hashes after generation but do not set video duration or aspect-ratio parameters.
  2. Save the generated video file name and dimensions after inference but never configure the model's supported duration or aspect-ratio controls.
  3. Apply supported video duration, aspect-ratio and resolution controls where available
  4. Retry a prior safety refusal rather than configuring supported video duration and aspect-ratio settings.

Explanation

Correct Answer

C:

Generation controls should be chosen from the model's actual supported capabilities and validated against delivery constraints. This resolves the specific need to control supported video presentation parameters, while a measure that helps trace synthetic imagery back to generating inputs addresses a different requirement.

Incorrect Answers

A:

Audit identifiers cannot control the output dimensions. Set the supported duration and aspect-ratio or size parameters and validate the resulting media against the brief.

B:

Recording file metadata is observational, not a generation constraint. Select the model's available duration and frame-size controls before generating and validate the returned clip.

D:

Content-policy retries do not correct a valid generated clip with the wrong presentation dimensions. Configure supported output controls for the target format.

Question 29

An auditor needs a short overview of a lengthy contract but cannot tolerate invented clauses. What summarization method fits?

  1. Generate a brief answer from the contract title alone, without reading clauses or retaining page references.
  2. Generate a source-grounded summary with cited contract sections
  3. Redact or mask sensitive fields before publishing the summary
  4. Retrieve relevant sections across the whole source before summarization

Explanation

Correct Answer

B:

Grounding with passage references helps reviewers verify that reported obligations occur in the document. Here the relevant capability must let the team summarize a long contract with traceable evidence. An alternative targeting preserve speaker commitments in conversation summaries solves another problem.

Incorrect Answers

A:

A title does not establish the actual contractual obligations, exceptions or limitations. Source-grounded summarization must use the contract clauses and preserve traceable evidence rather than inventing provisions.

C:

Data minimization must apply to summarization outputs as well as original inputs. That capability is appropriate for remove personal data from a shareable summary, not this separate task of finding how to summarize a long contract with traceable evidence.

D:

A summary cannot include critical evidence that prompt assembly never retrieved. Here the needed outcome is to summarize a long contract with traceable evidence; an option designed to include important evidence located outside opening sections would not directly provide it.

Question 30

A search index has empty content for scans of signed paper forms. Which enrichment should run first?

  1. Preserve OCR table-cell structure and bounding regions during extraction
  2. Run semantic reranking on the empty searchable text field of the scanned pages without first recognizing the characters.
  3. Use layout-aware reading order and column segmentation
  4. 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.

Developing AI Apps and Agents on Azure (AI-103) Practice Tests