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 64

Question 630

A policy-answering app must cite a document clause that changed during the current month. The search index includes document_id, revision and chunk_id fields. What should the response pipeline use as its citation?

  1. Use the retrieved source identifiers and approved revision rather than model-invented page numbers.
  2. Ask the generator to create an attractive citation number without reading search metadata.
  3. Cite the newest document title even when the retrieved chunk came from an older revision.
  4. Keep only the embedding distance and convert it into a clause number.

Explanation

Correct Answer

A:

Index metadata ties each claim to the actual document version used as evidence.

Incorrect Answers

B:

A formatted reference can look convincing while failing to resolve to any source.

C:

This misattributes old content to a different policy version and can conceal a change.

D:

Similarity scores cannot serve as stable identifiers for policy clauses.

Question 631

An AI sales assistant calls a stock service and receives a response containing available_quantity, reserved_quantity and timestamp. Users need the sellable number at that timestamp. Which method is best?

  1. Let the language model infer a likely stock level from product popularity trends.
  2. Use the latest conversational estimate instead of the tool result when they disagree.
  3. Interpret the service timestamp as the quantity and ignore the reservation field.
  4. Calculate available minus reserved from validated tool fields and include the recorded timestamp.

Explanation

Correct Answer

D:

A deterministic computation grounded in the API response gives a traceable sellable count.

Incorrect Answers

A:

Popularity is not an authoritative measurement of inventory available for sale.

B:

An earlier estimate is stale and should not override a current authoritative response.

C:

Time metadata cannot replace the quantity calculation required by the scenario.

Question 632

A procurement interface lets users preview a proposed purchase order before it is submitted. The model can return typed create_order arguments, but no purchase may occur until confirmation. How should the client behave?

  1. Execute every proposed tool call automatically and ask for confirmation afterward.
  2. Render validated proposed arguments for review and execute the tool only after separate confirmation.
  3. Translate the proposed arguments to plain text and skip business validation before execution.
  4. Treat a high model confidence score as a substitute for the user confirmation.

Explanation

Correct Answer

B:

Generating a tool-call proposal is distinct from authorizing a real purchasing side effect.

Incorrect Answers

A:

Post-execution approval cannot prevent an unauthorized order from being created.

C:

Presentation formatting does not validate item, price or permission constraints.

D:

Prediction confidence cannot grant authorization for a purchase.

Question 633

A RAG chatbot repeatedly retrieves the same disclaimer from ten related pages, leaving too little room for the actual answer. What prompt-construction change addresses the wasted context?

  1. Increase top_k to retrieve additional copies of the same disclaimer.
  2. Discard all retrieved passages and rely only on the assistant’s previous response.
  3. Deduplicate overlapping retrieved boilerplate while retaining distinct source evidence and citations.
  4. Remove source IDs from the prompt but keep every repeated disclaimer unchanged.

Explanation

Correct Answer

C:

Reducing repeated content preserves the token budget for relevant facts without losing provenance.

Incorrect Answers

A:

More duplicates increase token consumption and leave less capacity for the actual answer.

B:

Removing current evidence breaks the application’s grounding requirement.

D:

Deleting citations does not address redundant content and harms traceability.

Question 634

An employee benefits agent uses a shared search index. HR rules allow the employee to read only their own entitlement records, while policy documents are shared. What retrieval behavior is required?

  1. Apply caller-authorized record filters before passing private passages to the model.
  2. Retrieve all employee records and ask the model to avoid mentioning unauthorized ones.
  3. Rely solely on embedding similarity, since similar records are not identifiable.
  4. Include the employee name in a prompt and disable record-level authorization filters.

Explanation

Correct Answer

A:

The search layer must enforce the employee’s data scope independently of generation instructions.

Incorrect Answers

B:

Exposing private chunks to the model violates the least-privilege access boundary.

C:

Vector similarity does not enforce ownership or entitlement permissions.

D:

A name supplied in text is not a secure access-control decision.

Question 635

A contract states “the supplier shall not subcontract without approval.” A summary generated for legal review omits “not.” Which evaluation condition should the application require?

  1. Judge the summary solely by its fluency and average sentence length.
  2. Increase creativity to encourage a more natural interpretation of the prohibition.
  3. Use a semantic search score as proof that every summarized legal obligation is accurate.
  4. Check critical negations and obligations against cited original clauses before accepting the summary.

Explanation

Correct Answer

D:

Preserving prohibition wording is essential to avoid reversing the legal effect of the text.

Incorrect Answers

A:

Well-written prose can still invert a mandatory condition.

B:

Sampling variability provides no assurance that legal negation will be preserved.

C:

Retrieval relevance does not independently validate the truth of a generated summary.

Question 636

A search app changes its query embedding model but leaves existing indexed vectors from a model with a different dimensionality. Similarity queries begin failing. What must the migration address?

  1. Increase the language model maximum completion tokens without changing the vectors.
  2. Regenerate or migrate indexed embeddings and align vector field dimensions with the query model.
  3. Force the search engine to interpret mismatched vector lengths as the same representation.
  4. Disable retrieval diagnostics while retaining the incompatible index entries.

Explanation

Correct Answer

B:

Search vectors must be compatible with the model and dimensionality used for query generation.

Incorrect Answers

A:

Generation output limits are unrelated to vector dimension compatibility.

C:

Different vector shapes cannot be safely compared as if they were identical.

D:

Suppressing errors does not make the vector index queryable.

Question 637

A support agent compresses older chat history to stay within context limits. The summary says the user preferred email, but the latest turn explicitly requests a phone callback. Which instruction should the agent honor?

  1. Treat the compacted history as a higher-priority system rule because it is shorter.
  2. Ignore the latest turn because it arrived after the summary was generated.
  3. Prioritize the current authorized request while retaining older history as background context.
  4. Make both contact methods mandatory without seeking clarification or applying policy.

Explanation

Correct Answer

C:

A summary of previous preferences should not override a newer explicit user instruction.

Incorrect Answers

A:

Compression does not elevate previous conversational content to trusted authority.

B:

Recency alone does not invalidate an authorized change of preference.

D:

This can violate the explicit current preference and add an unrequested contact channel.

Question 638

A knowledge assistant produces a response with three numbered citations, but one number is not present in the retrieval results. The service must not show fabricated sources. What should the postprocessor do?

  1. Verify each citation against retrieved source IDs and remove or flag unsupported claims.
  2. Keep all citations because a language model assigned high confidence to the response.
  3. Convert the missing citation into the nearest numeric ID without checking its content.
  4. Remove the unsupported citation marker but continue stating its dependent claim as verified.

Explanation

Correct Answer

A:

Source-ID validation prevents presenting nonexistent evidence as an authentic document reference.

Incorrect Answers

B:

Confidence does not establish that a cited source exists in the retrieved set.

C:

An unrelated source number can misattribute the generated claim.

D:

Deleting the marker leaves an unsupported claim presented as established fact.

Question 639

An assistant answers English product queries accurately but fails on equivalent Spanish questions. Logs show that it retrieves no passages for Spanish queries. What should engineers test first?

  1. Increase completion temperature to make the Spanish responses more imaginative.
  2. Remove the retrieval tool and answer Spanish requests directly from model memory.
  3. Change the HTML page language tag without testing retrieval behavior.
  4. Measure cross-language query embedding and retrieval recall against the multilingual corpus.

Explanation

Correct Answer

D:

Empty evidence points to retrieval or language coverage rather than the answer generator itself.

Incorrect Answers

A:

More creative output cannot recover missing authorized source passages.

B:

Dropping evidence undermines accuracy for current product-specific information.

C:

Page markup does not establish multilingual search or embedding compatibility.

Question 640

A team introduces few-shot examples to improve invoice categorization. It has high-quality production invoices, but they contain real account numbers and names. What is a defensible prompt design?

  1. Place unredacted invoices in a globally shared system prompt for all tenants.
  2. Use minimized or synthetic representative examples with the same classification structure.
  3. Disable all classification examples and accept malformed output rather than sanitize data.
  4. Encode customer account numbers as visible text and rely on model secrecy to protect them.

Explanation

Correct Answer

B:

Representative examples can teach output patterns without exposing unnecessary customer identifiers.

Incorrect Answers

A:

This creates cross-tenant personal-data exposure outside the example’s teaching purpose.

C:

Data minimization allows useful examples while meeting privacy requirements.

D:

Encoding text does not replace explicit control over sensitive example data.

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