McAfee-Secured Website

Exam Code: AI-103

Exam Name: Developing AI Apps and Agents on Azure

Certification Provider: Microsoft

Corresponding Certification: Microsoft Certified: Azure AI Apps and Agents Developer Associate

100% Microsoft AI-103 Guaranteed Success With Testking Training Materials

Get Real AI-103 Exam Questions & Expert Verified Answers as Experienced on the Real Exam!

187 Questions & Answers with Testing Engine

"Developing AI Apps and Agents on Azure Exam", also known as AI-103 exam, is a Microsoft certification exam.

Pass your tests with the always up-to-date AI-103 Exam Engine. Your AI-103 training materials keep you at the head of the pack!

Satisfaction Guaranteed

Money Back Guarantee

Testking provides hassle-free money back guarantee with our products. That is because we have 100% trust in the abilities of our professional and experience product team, and our record is a proof of that.

99.6% PASS RATE
Was: $137.49
Now: $124.99

Microsoft AI-103 Practice Test 24

Question 230

A corporate policy changed last week, after the model's training data cutoff. How can the assistant answer the new rule accurately?

  1. Assume the model's general training contains the unpublished policy amendment.
  2. Use only last year's cached answer because it has a high confidence score.
  3. Use grounding data to answer policy changes newer than training cutoff.
  4. Increase response creativity and rely on a pretrained model instead of retrieving the revised internal policy.

Explanation

Correct Answer

C:

Fresh external evidence supplies changes absent from static training.

Incorrect Answers

A:

Pretraining cannot guarantee knowledge of a later private revision.

B:

Confidence in an old answer does not establish its current validity.

D:

Sampling variation cannot substitute for new, authoritative organizational source material.

Question 231

An HR user lost access to a sensitive folder yesterday, but a RAG index still contains the document. What must happen before source passages are displayed?

  1. Apply source ACLs again when retrieving/rendering citations.
  2. Allow any citation URL because source documents reside in the same index.
  3. Keep using the employee's old cached access list indefinitely.
  4. Rely on the model to hide the text after it has already received the restricted document.

Explanation

Correct Answer

A:

Authorization can change independently of indexing and must be enforced for the active caller.

Incorrect Answers

B:

Index membership does not confer per-user authorization.

C:

Stale entitlements can permit access after revocation.

D:

Post-retrieval concealment does not prevent exposure in model context.

Question 232

A diagnostics pipeline captures complete model prompts and all Azure access tokens to aid troubleshooting. What should be changed?

  1. Remove authentication from the service so there are no tokens to log.
  2. Retain plaintext credentials because only the operations team views logs.
  3. Send every captured prompt to an external public paste service.
  4. Redact credentials from model diagnostic logs.

Explanation

Correct Answer

D:

Sanitized diagnostic logging supports root-cause analysis without creating a token disclosure store.

Incorrect Answers

A:

Disabling identity is not acceptable protection for diagnostic data.

B:

Operations logs can be widely accessible or copied and should not hold reusable secrets.

C:

Public data transfer increases confidentiality risk and is unnecessary for troubleshooting.

Question 233

A RAG evaluation currently scores polished writing, but users report citations pointing to unrelated pages. Which benchmark should be added?

  1. Measure only model response latency.
  2. Score the answer's readability and citation format without validating the linked passages.
  3. Validate each cited claim against actual source passages.
  4. Count sentence length and treat shorter claims as better-cited.

Explanation

Correct Answer

C:

Citation accuracy requires matching claims to supporting evidence rather than judging prose style.

Incorrect Answers

A:

Fast output can still contain false citations.

B:

Readable prose and plausible citation formatting can conceal that cited passages don't substantiate the claims.

D:

Length does not prove source support.

Question 234

A customer-support bot serves several tenants and shares one backend process. A memory bug inserts a previous customer's issue into another customer's prompt. Which design boundary is essential?

  1. Keep a single global conversation buffer and ask the model not to mention other users.
  2. Scrub only the final UI after generation while keeping mixed histories in prompt context.
  3. Avoid cross-user conversation history leakage in a multi-tenant chat app.
  4. Assign all sessions the same conversation ID for cache efficiency.

Explanation

Correct Answer

C:

Per-principal state selection prevents one customer's context entering another's generation request.

Incorrect Answers

A:

The sensitive content has already crossed the isolation boundary.

B:

Post-generation masking cannot undo cross-user data exposure to the model.

D:

Shared identity keys increase the risk of cross-session history collisions.

Question 235

A handbook RAG assistant answers questions even when the corpus has no relevant support. What negative cases should evaluation include?

  1. Remove all questions with empty search results from the evaluation set.
  2. Treat any plausible answer as correct when the model has high confidence.
  3. Use deliberately unanswerable questions and require evidence-insufficient behavior without fabricated claims or citations.
  4. Ask only questions copied word-for-word from existing handbook titles.

Explanation

Correct Answer

C:

Negative test cases measure whether the assistant recognizes lack of grounding.

Incorrect Answers

A:

Excluding failures conceals hallucination behavior outside corpus coverage.

B:

Confidence cannot create missing source evidence.

D:

Easy matching cases do not exercise absence-aware decisions.

Question 236

An index returns five copies of the same paragraph from overlapping documents, crowding out diverse relevant context. Which prompt assembly improvement is useful?

  1. Use the highest-scoring single retrieval hit without checking relevance or whether the answer needs other clauses.
  2. Remove all citations so duplicate evidence cannot be noticed.
  3. Increase top-k until the same paragraph appears twenty times.
  4. Implement context deduplication to prevent repeated retrieved passages.

Explanation

Correct Answer

D:

Removing redundant contexts frees tokens for varied supporting source information.

Incorrect Answers

A:

Top-1 retrieval can omit relevant exceptions or depend on irrelevant matches.

B:

Suppressing source metadata obscures rather than fixes retrieval redundancy.

C:

Additional copies waste context and can bias generation.

Question 237

An agent drafts a vendor change request but policy requires an employee to approve before any external submission. What architectural separation should exist?

  1. Record an approval only after the API confirms the vendor change.
  2. Send the drafted request immediately whenever the model says 'ready'.
  3. Grant the agent a global service credential and expect a prompt to enforce review.
  4. Separate draft preparation and review from authorized external submission.

Explanation

Correct Answer

D:

Independent approval before the side effect preserves control over external transactions.

Incorrect Answers

A:

Post-action logging does not meet the pre-submission approval rule.

B:

An LLM's wording is not the required approval signal.

C:

Text instructions cannot independently authorize a business submission.

Question 238

A model writes SQL filters for a report API. One generated condition includes an unknown column and an injection-like operator. What must happen before query execution?

  1. Concatenate the raw model output into a privileged SQL query.
  2. Assign the model full database-owner rights so unsupported columns can be created.
  3. Add application-side validation to model-generated SQL filter expressions.
  4. Check only whether the generated SQL contains the word SELECT.

Explanation

Correct Answer

C:

A constrained query builder prevents arbitrary model strings from becoming executable database commands.

Incorrect Answers

A:

Direct string execution creates syntax, integrity and injection risks.

B:

Privilege expansion worsens risks and does not solve validation.

D:

One keyword is not a safety or schema validation of a complete query.

Question 239

A FAQ assistant sometimes calls an external CRM for every greeting, creating latency and privacy exposure even when no customer data is needed. What routing change helps?

  1. Keep external requests out of the critical answer path when not required.
  2. Call every connected service preemptively for all user messages.
  3. Require a full database scan before answering 'hello'.
  4. Cache CRM access tokens in the user-visible conversation to save setup time.

Explanation

Correct Answer

A:

Avoiding unnecessary dependencies improves latency and data minimization.

Incorrect Answers

B:

Unneeded requests add cost, failures and unnecessary data access.

C:

A greeting has no dependency on live business records.

D:

Exposing credentials is not a safe performance optimization.

Question 240

A model answer is truncated even at low temperature because the API has a small `max_output_tokens` value. Which parameter should be adjusted?

  1. Raise search top-k without changing the completion limit.
  2. Increase temperature while leaving the same tiny output cap.
  3. Increase the output-token limit within model and cost constraints.
  4. Disable source citations because the model will gain unlimited completion capacity.

Explanation

Correct Answer

C:

The completion length cap directly limits emitted tokens, unlike sampling parameters.

Incorrect Answers

A:

Retrieval breadth does not enlarge the response budget.

B:

Sampling variability cannot extend a hard output-token allowance.

D:

Removing references does not eliminate the explicit token cap.

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