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 10

Question 90

A search result has a matching heading but its body discusses a different product. What should the grounding layer do?

  1. Accept the heading match as sufficient evidence and cite the snippet without examining the body text or product identifier.
  2. Filter or rerank using substantive passage content rather than title matches alone
  3. Highlight the documented conflict and route the issue to the authorized policy owner
  4. Attach source version and retrieval timestamp to citations

Explanation

Correct Answer

B:

The useful evidence must actually answer the question, not merely contain similar headline words. Here the relevant capability must let the team reject superficial result-title matches. An alternative targeting compute correct numbers from retrieved financial data solves another problem.

Incorrect Answers

A:

Headings can be duplicated across product families. Grounding should inspect the substantive matched passage and its product context before accepting it as evidence for the user's request.

C:

Conflicting authoritative passages require resolution, not silent selection of whichever rank is highest. That capability is appropriate for reconcile contradictory retrieved policies transparently, not this separate task of finding how to reject superficial result-title matches.

D:

Version provenance establishes whether the answer relied on current or historical reference material. Here the needed outcome is to reject superficial result-title matches; an option designed to identify the precise evidence version would not directly provide it.

Question 91

Company documents are updated nightly yet an agent cites outdated versions. Which pipeline behavior should be implemented?

  1. Rebuild vector embeddings for unchanged documents but never update or remove entries whose source policy versions changed overnight.
  2. Rerank the outdated index nightly instead of reconciling newly approved and deleted source revisions.
  3. Increase keyword match weight on obsolete document titles and leave old indexed chunks active after nightly source updates.
  4. Schedule incremental ingestion with update and deletion tracking

Explanation

Correct Answer

D:

Index freshness depends on propagating changes and removals from authoritative sources into searchable content. In this scenario the team must keep indexed content synchronized with updates. A feature used to search exact terms alongside semantic similarity does not substitute for it.

Incorrect Answers

A:

Re-embedding stale content leaves the same obsolete facts searchable. Incremental ingestion must track changed, added and deleted source revisions.

B:

Semantic ranking cannot make absent updates or deleted documents current. Refresh and reconcile the index from authoritative revisions before sorting candidate results.

C:

Keyword scoring cannot establish which revision is currently effective. Index update and deletion reconciliation must remove or demote superseded source versions.

Question 92

A grounded application has conflicting policy documents and a weekly inventory report while a user requests a current quantity. Which TWO actions avoid unsupported certainty? Choose TWO.

  1. Average the weekly inventory quantity with the two conflicting policy passages and report the result as today's authoritative stock count.
  2. State that the available sources do not establish the requested procedure
  3. Call the authoritative live inventory system for real-time figures
  4. Calculate the current inventory amount from last week's report as though no transactions had occurred since that snapshot.
  5. Highlight the documented conflict and route the issue to the authorized policy owner

Explanation

Correct Answers

C:

A periodically updated index is not a reliable source for a fact that changes minute by minute. Escalating unresolved policy conflict and querying live inventory each resolve a distinct evidence-quality limitation.

E:

Conflicting authoritative passages require resolution, not silent selection of whichever rank is highest.

Incorrect Answers

A:

Arithmetic over stale inventory and contradictory policy text cannot establish live availability. Query the authoritative live system and disclose unresolved policy conflicts.

B:

Explicit uncertainty is preferable to fabricating an internal policy not present in accessible evidence. That is useful for handle missing evidence without making up facts, but it does not directly accomplish the requested need to resolve conflicting policy evidence and obtain real-time quantities.

D:

Arithmetic over a stale snapshot cannot establish a real-time inventory balance. The live inventory system is required, while conflicting policies need explicit owner resolution.

Question 93

A compliance team must cite precise text found in a scanned PDF that includes tables. What indexing enrichment is appropriate?

  1. Invoke semantic ranking over a scanned PDF whose table text was never extracted into the index.
  2. Generate embeddings directly from scanned page bitmaps while omitting OCR character recognition and table structure needed for exact quotations.
  3. Chunk source documents into coherent, indexed passages with source offsets
  4. Apply OCR and layout-aware extraction before indexing

Explanation

Correct Answer

D:

OCR and layout processing recover text and table relationships from images that ordinary text extraction may miss. This resolves the specific need to index scanned text while preserving table context, while a measure that helps rerank relevant candidates for question fit addresses a different requirement.

Incorrect Answers

A:

The reranker works with searchable text in candidates and cannot read missing scan content. Run appropriate OCR and layout/table extraction before indexing and citation.

B:

Visual embeddings alone do not provide the required verbatim clause and cell contents. OCR plus layout extraction must recover searchable text with page-level provenance.

C:

Passage granularity improves retrieval specificity while source offsets preserve provenance for citations. This helps when a team needs to retrieve useful subsections of long documents; it does not establish how to index scanned text while preserving table context.

Question 94

An application streams an answer and then discovers it contains a prohibited tool action. How should the execution boundary be redesigned?

  1. Execute tool calls as soon as their arguments appear in streamed model text.
  2. Apply content moderation only to the final displayed answer after the tool has run.
  3. Run required tool authorization and policy checks before permitting the side-effecting call.
  4. Trust the model to label its own tool actions as approved.

Explanation

Correct Answer

C:

Streaming a response does not grant permission to execute actions; checks must occur at the trusted controller and tool backend.

Incorrect Answers

A:

Executing an unvalidated partial request can cause prohibited side effects before policy checks finish.

B:

A final-answer moderation check happens too late when a forbidden tool action has already changed an external system. Enforce policy before the tool call and keep subsequent display review as an additional safeguard.

D:

Self-asserted approval is not a substitute for server-side authorization at the action boundary.

Question 95

A support RAG solution must find both exact product serials and paraphrased symptoms, while keeping document revisions current. Which THREE retrieval measures address those separate needs? Choose THREE.

  1. Schedule incremental ingestion with update and deletion tracking
  2. Retain lexical keyword retrieval for exact part-number matching
  3. Rely exclusively on a static semantic index, ignoring exact product serial tokens and document updates until a manual rebuild occurs.
  4. Use only dense vector similarity and omit exact-identifier matching and refresh policies.
  5. Add semantic reranking alone but omit both exact-code keyword recall and source-revision synchronization.
  6. Combine lexical and vector queries in hybrid search

Explanation

Correct Answers

A:

Index freshness depends on propagating changes and removals from authoritative sources into searchable content. Keyword capability preserves exact serials, hybrid search supports the mixed query, and incremental refresh prevents stale material.

B:

Token/keyword search can preserve identifier matching that purely semantic similarity might weaken.

F:

Hybrid retrieval can balance exact-code precision with semantic symptom recall in a single retrieval plan.

Incorrect Answers

C:

A single stale vector-only index misses exact identifiers and changing facts. Lexical-plus-vector retrieval must be combined with an update/deletion ingestion process.

D:

Dense-only search would fail the explicit exact-serial and freshness requirements, so it cannot replace the hybrid retrieval and incremental update steps.

E:

A reranker improves order only for already retrieved records. The solution still requires hybrid code-and-symptom retrieval and removal or replacement of superseded documents.

Question 96

A Claude-style agent proposes a database-write tool call but the request fails schema validation. What must happen before execution?

  1. Execute the account-changing function with guessed default fields when its argument object fails required schema validation.
  2. Parallel-execute a database write before its tool arguments pass the required input schema.
  3. Reject invalid tool arguments and request bounded correction before any side effect
  4. Wait for an unrelated balance lookup and then execute a database write with the same invalid argument schema.

Explanation

Correct Answer

C:

A model tool call must satisfy deterministic schema validation and business rules before a sensitive write. The decisive difference is whether the design can block malformed tool write arguments; the competing function serves to avoid duplicate writes after timeout.

Incorrect Answers

A:

Guessing values for a write tool risks changing the wrong record. Reject the invalid request and obtain a bounded corrected payload before any side effect.

B:

Concurrency doesn't repair malformed arguments. Reject or correct the proposed write before any execution and enforce the action's schema and authorization boundary.

D:

Waiting for a prerequisite cannot make structurally invalid arguments valid. Reject the write and obtain a schema-conforming, authorized request before execution.

Question 97

Two independent data lookups have no dependency, yet the controller performs them sequentially. What may improve latency?

  1. Run independent tool calls concurrently and correlate their results to the correct requests
  2. Duplicate each independent lookup across extra sequential agent turns and merge results without request IDs.
  3. Serialize independent account and inventory lookups even though neither result is required to construct the other.
  4. Reject invalid tool arguments and request bounded correction before any side effect

Explanation

Correct Answer

A:

Parallelism is appropriate when neither result determines the other call's inputs. In this scenario the team must parallelize independent retrieval work. A feature used to stop runaway tool usage does not substitute for it.

Incorrect Answers

B:

Additional sequential turns increase latency and can mix responses. Parallel independent calls with stable correlation identifiers are appropriate when there is no dependency.

C:

The calls are unrelated reads and can be run concurrently with correct request/result correlation. Unnecessary dependency adds latency without improving correctness.

D:

A model tool call must satisfy deterministic schema validation and business rules before a sensitive write. This helps when a team needs to block malformed tool write arguments; it does not establish how to parallelize independent retrieval work.

Question 98

A refund tool's parameter depends on an earlier balance lookup. How should the controller schedule calls?

  1. Run the two independent lookups serially and discard the request IDs that would correlate their responses.
  2. Construct the refund amount from the user's original estimate before the required balance lookup returns.
  3. Wait for the balance result before constructing the dependent refund request
  4. Limit tool iterations and elapsed time and detect repeated no-progress calls

Explanation

Correct Answer

C:

Dependent actions cannot be configured correctly before their prerequisite data arrives. This resolves the specific need to respect prerequisite tool result order, while a measure that helps parallelize independent retrieval work addresses a different requirement.

Incorrect Answers

A:

Serial calls add avoidable latency to independent work, while lost correlation can mix results. Parallelize eligible reads and associate every result with its originating request.

B:

The refund is dependent on verified balance data. Executing first would risk an inaccurate or unauthorized amount; wait for the lookup before preparing the write.

D:

An external execution budget keeps autonomy bounded even if prompts would otherwise loop. The function serves to stop runaway tool usage, whereas the actual requirement is to respect prerequisite tool result order.

Question 99

A Python client expects a JSON invoice object but receives explanatory prose before the JSON. What output mechanism should be considered?

  1. Strip all characters before the first opening brace and trust the resulting JSON.
  2. Set temperature to zero and assume prose wrappers will never appear.
  3. Parse the entire free-form response as a dictionary without checking types.
  4. Use the supported structured-output or tool-schema interface and validate the returned fields.

Explanation

Correct Answer

D:

Schema-constrained output helps reduce formatting drift; business rules and unexpected refusals still need application-side validation.

Incorrect Answers

A:

Trimming prose is brittle and does not validate required fields or handle malformed and truncated output.

B:

A low temperature does not guarantee strict schema compliance or eliminate refusal and truncation cases.

C:

A parser failure or unexpected field type can corrupt downstream invoice processing.

Question 100

A payment provider returns a stable transaction reference but does not support idempotency keys. After a timeout, the assistant cannot tell whether the charge completed. What recovery step should run before attempting any second charge?

  1. Issue a second card charge immediately after the timeout because the request did not return a success message.
  2. Construct a refund immediately without reconciling whether the original timed-out charge was completed.
  3. Query the provider for the result associated with the original transaction reference and retry only after confirming no charge was posted.
  4. Start a second charge in parallel before checking whether the first timed-out charge already committed.

Explanation

Correct Answer

C:

When the provider has no idempotency-key support, a client should query the authoritative transaction status using the original reference before issuing another side-effecting request. A timed-out HTTP response does not prove the charge failed. Reconciliation, not blind retry, is the decisive step. Read-before-retry verification uses the authoritative transaction reference to determine whether the first write succeeded, avoiding a second charge when the initial response was lost.

Incorrect Answers

A:

An uncertain timeout does not prove the first charge failed. Query the provider using the original transaction reference to avoid duplicate billing before considering a retry.

B:

A refund is a separate side effect and may be wrong if no charge occurred. Query the provider using the transaction reference before deciding an authorized remedy.

D:

An ambiguous payment timeout is not evidence of failure. Blind concurrent retries may double-charge; reconcile the existing transaction reference with the provider first.

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