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 13

Question 120

A retrieved video snippet must be reviewed in the original file. What provenance should the index store?

  1. Save only a generated clip caption without the original video asset ID or time interval
  2. Associate each scene hit with video ID and precise time range
  3. Store the scene's textual caption without a video identifier or start and end time markers linking it to the original recording.
  4. Extract timestamped scenes or frames and generate indexed visual descriptors

Explanation

Correct Answer

B:

Timestamped source links let reviewers validate a visual-search claim. This directly supports the need to link scene results to original footage, unlike a mechanism primarily intended to search image collections by visual similarity.

Incorrect Answers

A:

A caption is not a reliable pointer to the evidence location. Video retrieval results need an asset identifier and exact timestamps so a reviewer can replay the claimed event.

C:

A caption alone cannot navigate a reviewer to the precise evidence. The index needs the source video ID and time range for each retrieved scene.

D:

Segment-based visual metadata makes otherwise opaque video content searchable. Its primary purpose is to create searchable video-scene metadata; the stated constraint instead demands a way to link scene results to original footage.

Question 121

A financial agent performs side-effecting tool calls and must recover from invalid arguments, uncertain timeouts and tool failures. Which THREE controller behaviors prevent silent damage? Choose THREE.

  1. Treat every financial tool as dependent on a balance query and skip argument validation or payment-timeout reconciliation.
  2. Reject invalid tool arguments and request bounded correction before any side effect
  3. After a failed write, suppress the tool error and retry with newly guessed arguments, without checking whether a prior attempt posted.
  4. Use a stable idempotency identifier and reconcile the existing invoice before retry
  5. Replay all uncertain financial writes concurrently without first resolving their transaction identifiers.
  6. Return a structured error result preserving the failed call's status

Explanation

Correct Answers

B:

A model tool call must satisfy deterministic schema validation and business rules before a sensitive write. Validation before writes, idempotent reconciliation and explicit error propagation address independent failure modes.

D:

Uncertain completion can duplicate financial side effects; idempotency and read-before-retry reduce that risk.

F:

A failed lookup is not the same as an empty result; the agent should be able to distinguish and recover.

Incorrect Answers

A:

Dependency scheduling is only one control. Invalid arguments need rejection, uncertain charges need provider reconciliation, and actual tool failures must remain visible.

C:

A blind retry can duplicate financial side effects. The controller must validate arguments, preserve errors, and reconcile uncertain outcomes with stable identifiers.

E:

Repeating side-effecting operations is unsafe when the provider's previous result is unknown. Validate arguments, reconcile uncertain writes and use bounded retries with clear failure reporting.

Question 122

A multi-user agent accidentally recalls another employee's prior private conversation. What state constraint addresses it?

  1. Store all employees' conversation memory under one shared cache key and rely on the assistant to avoid revealing it.
  2. Summarize or prune stale conversation state while retaining verified task constraints
  3. Store explicit durable preferences in a governed profile store, excluding secrets
  4. Partition conversation memory by authenticated user and authorization context

Explanation

Correct Answer

D:

User-scoped state prevents retrieval of unrelated private conversations when separate users share a model deployment. In this scenario the team must isolate memory between users. A feature used to filter retrieved knowledge by the user's permission does not substitute for it.

Incorrect Answers

A:

A shared memory key breaks tenant and user isolation before the model sees the data. Partition access by authenticated principal and enforce authorization before retrieval.

B:

Selective context reduces token bloat and avoids treating old statements as current authoritative instructions. Its primary purpose is to maintain compact current context; the stated constraint instead demands a way to isolate memory between users.

C:

Long-lived user preferences need a controlled source with consent and deletion policies rather than raw unbounded chat logs. That is relevant to persist allowed user preferences across sessions, but the present scenario requires an ability to isolate memory between users.

Question 123

A product photo editor must replace one item without changing nearby labels and later prove what area was edited. Which THREE practices support that? Choose THREE.

  1. Retain source media, mask versions, edit prompt and output version IDs
  2. Regenerate every video frame from scratch to replace a small product photo object, without preserving the image mask or source asset.
  3. Use mask-based image inpainting on the marked region
  4. Evaluate preservation of unmasked regions as a specific acceptance criterion
  5. Apply the mask without checking source-image resolution, orientation or selected pixels.
  6. Run a temporal video edit on the product photo and omit both the label-preserving mask and its source lineage.

Explanation

Correct Answers

A:

Versioning masks and source assets makes region editing auditable and reproducible. Mask-based inpainting, preservation checks and saved edit provenance jointly deliver controlled editing.

C:

Inpainting targets a defined spatial area instead of regenerating an entire image unnecessarily.

D:

Localized edits should be tested for unintended changes outside the approved region.

Incorrect Answers

B:

Unrestricted regeneration cannot confine an edit or document which area changed. The photo workflow requires a versioned mask, protected-region verification and recorded provenance.

E:

A misregistered mask may alter nearby labels even when the editing tool otherwise supports inpainting.

F:

Multi-frame tools do not replace still-image region protection. The requirement combines a constrained edit, preservation of nearby labels and reproducible mask/source evidence.

Question 124

A contract assistant returns JSON where a date uses two different locale formats across runs. What contract should be enforced?

  1. Allow either local date order and rely on downstream users to infer it.
  2. Ask for a human-readable date sentence rather than a typed field.
  3. Require a specified ISO-style date field schema and validate the parsed value.
  4. Change decimal-separator settings but leave the date field unconstrained.

Explanation

Correct Answer

C:

Machine-readable output needs stable field formats and validation; valid-looking text may still be ambiguous to downstream parsers.

Incorrect Answers

A:

Ambiguous date strings can be interpreted differently across locales and cause silent processing errors.

B:

Prose does not impose a stable parseable date representation for JSON consumers.

D:

Changing a numeric decimal separator cannot standardize a date representation in JSON. Specify an agreed date format such as a valid ISO-style date string in the response schema and validate the parsed field.

Question 125

A RAG pipeline indexes document text but not references to figures that support the answer. What metadata should be captured?

  1. Index only OCR text and discard the figures and page geometry.
  2. Attach figure identifiers, page references and visual-region evidence to the indexed content.
  3. Attach generic image thumbnails without page or figure identifiers.
  4. Use the document title as the sole reference for every figure-based claim.

Explanation

Correct Answer

B:

Text-only chunks cannot cite visual evidence that was discarded during ingestion.

Incorrect Answers

A:

Claims supported by a figure cannot be verified if the figure and its location are not indexed.

C:

Unlinked thumbnails cannot map generated statements back to specific visual evidence.

D:

A title does not identify which diagram or page region supports the answer.

Question 126

A generated report stops midway through a table when the completion token cap is reached. Which signal should the client handle?

  1. Inspect the model's completion or stop status and retry with bounded context/output settings when truncated.
  2. Assume a successful HTTP status proves the model completed the entire table.
  3. Retry the same oversized request indefinitely without inspecting termination cause.
  4. Discard the incomplete table and silently present it as complete.

Explanation

Correct Answer

A:

Token exhaustion can yield incomplete output even when earlier tokens look well formed; downstream systems must not silently accept partial results.

Incorrect Answers

B:

Transport success does not prove completion; the generation can stop because its output budget was exhausted.

C:

Blind retries repeat the truncation and can increase usage without yielding a full report.

D:

Omitting the stop status hides material missing rows and fails the requirement for a full report.

Question 127

An assistant must remember user-approved preferences between separate sessions but avoid storing passwords. What pattern fits?

  1. Keep approved preferences only in an expiring session transcript and require the user to repeat them on every new login.
  2. Connect the agent to a permission-filtered retrieval knowledge source
  3. Summarize or prune stale conversation state while retaining verified task constraints
  4. Store explicit durable preferences in a governed profile store, excluding secrets

Explanation

Correct Answer

D:

Long-lived user preferences need a controlled source with consent and deletion policies rather than raw unbounded chat logs. This resolves the specific need to persist allowed user preferences across sessions, while a measure that helps isolate memory between users addresses a different requirement.

Incorrect Answers

A:

Session-only context is not durable across logins. Store consented nonsecret preferences in a governed profile service with retention and deletion controls.

B:

Knowledge retrieval supplies current citations on demand while avoiding huge prompts and unauthorized corpus exposure. That is relevant to ground answers in governed current documents, but the present scenario requires an ability to persist allowed user preferences across sessions.

C:

Selective context reduces token bloat and avoids treating old statements as current authoritative instructions. This helps when a team needs to maintain compact current context; it does not establish how to persist allowed user preferences across sessions.

Question 128

An image-generation app must reproduce a reference person's portrait supplied by a customer. What governance step comes before output?

  1. Regenerate the likeness without checking customer ownership of the reference photo.
  2. Replace the watermark and assume the portrait is cleared for any promotional use.
  3. Verify usage rights and consent for the reference likeness under the applicable policy.
  4. Approve reproduction solely because the generated portrait looks realistic.

Explanation

Correct Answer

C:

A reference image may contain a real person's likeness; permission and provenance should be checked before transformation.

Incorrect Answers

A:

Supplying an image does not automatically grant rights to reproduce a person's likeness.

B:

Removing a watermark does not establish consent, rights or permissible use of a reference person's image.

D:

Visual quality is independent of consent, licensing and applicable likeness safeguards.

Question 129

A message mixes praise of delivery with frustration about packaging. How should sentiment be assessed?

  1. Emit one document-level positive score and discard the negative packaging feedback
  2. Return one overall positive label based on praise for delivery, ignoring a separate negative judgment about packaging.
  3. Use contextual examples and evaluate sarcasm-specific errors
  4. Score sentiment at the aspect or topic level rather than only document level

Explanation

Correct Answer

D:

Aspect sentiment avoids collapsing mixed reactions into one misleading global score. This resolves the specific need to identify mixed sentiment about separate product aspects, while a measure that helps separate message intent from emotional tone addresses a different requirement.

Incorrect Answers

A:

One aggregate sentiment label conceals opposing opinions about different product aspects. Aspect-level evaluation must preserve delivery praise and packaging criticism separately.

B:

Document-level polarity hides mixed feedback. Aspect-level sentiment must score delivery and packaging separately to preserve the customer's different judgments.

C:

Literal polarity words can contradict contextual meaning, so domain testing is necessary. This helps when a team needs to recognize sentiment polarity in sarcastic context; it does not establish how to identify mixed sentiment about separate product aspects.

Question 130

A marketing repository includes image captions but no ability to find pictures by semantic visual similarity. What additional field helps?

  1. Index image file names as text and assume that alone supports semantic visual similarity
  2. Compute supported image embeddings and index them for vector search
  3. Index image filenames and their human-written captions only, without computing any embedding from the actual pixels.
  4. Preserve modality, asset and temporal relationships across indexed segments

Explanation

Correct Answer

B:

Visual embeddings can represent image content for semantic retrieval beyond caption terms. The decisive difference is whether the design can search image collections by visual similarity; the competing function serves to link scene results to original footage.

Incorrect Answers

A:

File names rarely encode visual content reliably. Images need an appropriate visual representation, such as supported embeddings, for semantic visual search.

C:

Text metadata cannot reliably support visual similarity between images with different captions. Visual embeddings are needed for nearest-neighbor image search.

D:

Typed links prevent text from one source segment being misattributed to another. That is relevant to maintain multimodal source associations, but the present scenario requires an ability to search image collections by visual similarity.

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