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 70

Question 690

A knowledge index receives the same approved policy PDF through a SharePoint export and a blob-storage mirror. Both copies create separate identical chunks, crowding retrieval results. What ingestion control should be introduced?

  1. Deduplicate on canonical document identity and revision while preserving intentional distinct versions.
  2. Index every mirror as an unrelated policy and let high top_k values absorb the duplicates.
  3. Delete all historical policy versions whenever any duplicate is detected.
  4. Ignore connector source identifiers and hash only the search query text.

Explanation

Correct Answer

A:

Source-aware deduplication removes redundant mirrors without losing legitimate revision history.

Incorrect Answers

B:

Repeated copies can crowd out other relevant evidence and waste context tokens.

C:

Different approved revisions may still be required for historical questions.

D:

A user query hash cannot establish whether two indexed chunks are the same document revision.

Question 691

A safety analytics agent searches incident narratives by semantic similarity but may return only incidents dated after 1 January 2026. The current index stores dates as unstructured text. What index and query change meets the rule?

  1. Add the cutoff to the generative system prompt and retrieve incidents from all years without filtering.
  2. Sort results by narrative similarity and assume the highest-scoring items are the newest.
  3. Apply the cutoff only to the first five returned items after limiting the candidate set.
  4. Store incident_date as a typed filterable field and apply the date predicate with the vector search.

Explanation

Correct Answer

D:

A metadata filter enforces the explicit cutoff while semantic retrieval ranks only eligible incidents.

Incorrect Answers

A:

The model may still receive ineligible records because prompt text does not enforce search eligibility.

B:

Similarity to the query has no necessary relationship to the incident date.

C:

Post hoc filtering can omit relevant eligible results that were not among the initially retrieved candidates.

Question 692

A contract search application often returns a clause but not the paragraph containing its exception. The exception begins immediately after a page break. What ingestion change should be evaluated?

  1. Split every page independently with no overlap because page boundaries imply separate legal meaning.
  2. Chunk by logical sections with measured overlap so dependent rule and exception remain available.
  3. Index only table-of-contents entries and discard full contract clauses.
  4. Increase model creativity until it infers an exception not present in the retrieved text.

Explanation

Correct Answer

B:

Document structure and overlap can preserve context separated by arbitrary page boundaries.

Incorrect Answers

A:

Page layout does not guarantee that a legal qualification is self-contained.

C:

Headings cannot provide the substantive rule and exception evidence.

D:

Generation cannot replace missing source passages in a grounded answer.

Question 693

An HR search corpus contains a policy approved today but not effective until next month. Staff ask for current leave entitlements. Which retrieval condition prevents citing the future rule as already applicable?

  1. Sort by upload time and cite the newest document without reading its effective date.
  2. Treat every approved policy as active immediately regardless of its stated commencement.
  3. Filter evidence by policy effective date relative to the question’s applicable date.
  4. Use embedding relevance to decide whether a policy date has arrived.

Explanation

Correct Answer

C:

Approval time and effective time can differ, so current-answer retrieval needs the valid interval.

Incorrect Answers

A:

The most recently uploaded rule may not yet be in force.

B:

Approval alone does not override the future-effective condition.

D:

Similarity ranking cannot replace comparison against an explicit effective date.

Question 694

An ingestion job reports two errors: a source server returned temporary HTTP 503, and a different document uses an unsupported file format. What failure-handling design is appropriate?

  1. Retry transient source requests with bounds, but route unsupported files to conversion or explicit error handling.
  2. Retry both errors indefinitely with identical settings until the daily quota is exhausted.
  3. Silently skip the failed documents while marking the entire source collection complete.
  4. Rebuild all successful indexed documents before inspecting the distinct error categories.

Explanation

Correct Answer

A:

Availability faults and format incompatibilities require different remediation paths.

Incorrect Answers

B:

A permanent unsupported format will not become valid through repeated identical requests.

C:

Unrecorded omissions make indexing completeness and search freshness impossible to assess.

D:

A full rebuild is unnecessary and may repeat the same permanent failure.

Question 695

A multinational invoice contains a subtotal in EUR and a separately stated total in USD. The payable system must not treat the two amounts as directly additive. What structured extraction is required?

  1. Combine the amounts as plain numbers and assign the company’s default currency.
  2. Convert all values by guessing an exchange rate from the document font and layout.
  3. Keep only the largest number and discard the other amount as duplicated billing.
  4. Extract each monetary value with its explicit currency and source location before reconciliation.

Explanation

Correct Answer

D:

Currencies and evidence offsets are needed to prevent mixing incompatible numeric amounts.

Incorrect Answers

A:

Ignoring units can create a materially incorrect financial total.

B:

Presentation details do not provide a trustworthy conversion rate.

C:

Different currencies and fields cannot be dismissed solely by magnitude.

Question 696

A document analyzer extracts subtotal 200, tax 20 and grand total 240 from a vendor invoice. The accounting workflow expects grand total to equal subtotal plus tax unless other charges are shown. What should it do?

  1. Post 240 because it is the extracted grand total regardless of the arithmetic conflict.
  2. Flag the inconsistency and verify source fields or additional fees before posting.
  3. Silently replace the subtotal with 220 without referencing the original invoice.
  4. Increase the model generation temperature to produce a more plausible total.

Explanation

Correct Answer

B:

A deterministic reconciliation reveals a twenty-unit discrepancy that requires investigation.

Incorrect Answers

A:

An extracted field can be wrong, and the stated totals do not reconcile on the available evidence.

C:

Inventing a new value corrupts source fidelity instead of resolving the discrepancy.

D:

Sampling settings cannot establish the missing financial evidence.

Question 697

Different suppliers label the same invoice identifier as “Invoice No.”, “Bill Ref.” and “Document ID.” An analyzer must populate one normalized invoice_number field. Which design supports that requirement?

  1. Create a separate payable database column for every label spelling found.
  2. Treat each label variant as a different invoice type and drop unmatched identifiers.
  3. Provide a normalized target field with descriptions and examples of legitimate label variants.
  4. Infer the invoice number from the upload filename whenever the visible document label differs.

Explanation

Correct Answer

C:

Semantic field guidance helps map diverse source labels to the same structured business value.

Incorrect Answers

A:

Source vocabulary variants do not require distinct business fields.

B:

Discarding valid aliases would create incomplete records.

D:

Filenames are not an authoritative substitute for an extracted source field.

Question 698

An onboarding form contains adjacent checkboxes labeled “paper billing” and “email billing.” The analyzer finds one checked box but loses its relationship to the labels. What extraction output is needed?

  1. Preserve each checkbox state together with its associated label and page region.
  2. Return only a count of checked boxes without the related field names.
  3. Assume the first label is selected whenever any checkbox on the page is marked.
  4. Convert all checkbox values to free-form positive sentiment labels.

Explanation

Correct Answer

A:

Spatial relationships allow downstream systems to distinguish the selected billing preference.

Incorrect Answers

B:

A count cannot establish which billing option the applicant selected.

C:

Page order alone does not prove the marked box’s actual association.

D:

Sentiment cannot represent the structured selection captured by the form.

Question 699

A compliance application must analyze 50,000 historical PDF files overnight and save structured fields for later review. Interactive response time is not required. Which processing design is appropriate?

  1. Open a separate synchronous browser chat for every PDF and discard job-status metadata.
  2. Skip failed files silently so the remaining files finish within the overnight window.
  3. Create summaries without retaining identifiers linking results to original PDF files.
  4. Use a supported asynchronous or batch analysis workflow with tracked job completion and errors.

Explanation

Correct Answer

D:

Bulk jobs can process the archive reliably while recording failures and results for audit.

Incorrect Answers

A:

Manual synchronous sessions do not provide an efficient, auditable bulk pipeline.

B:

Without error accounting, the claimed completeness of the archive cannot be verified.

C:

Unlinked outputs cannot support document-level reconciliation or review.

Question 700

An onboarding form contains an applicant name and postal address but no date of birth. A generative document analyzer fills birth_date using an unsupported guess. What output policy should be enforced?

  1. Accept the guessed birth date whenever the applicant’s other fields look plausible.
  2. Return an explicit missing or null birth_date and require source evidence for populated identity fields.
  3. Derive birth_date from postal code averages without disclosing the estimation.
  4. Populate birth_date with the document scan timestamp to satisfy the database.

Explanation

Correct Answer

B:

Extraction must represent the document faithfully rather than invent a sensitive personal attribute.

Incorrect Answers

A:

Plausibility is not source evidence for an absent identity attribute.

C:

Aggregate demographic information cannot establish one person’s date of birth.

D:

The scan date is unrelated to the applicant’s birth date and would corrupt the record.

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