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 69

Question 680

A finance assistant extracts a 2.75% fee and a 30-day deadline from an agreement. A prior version of the prompt changed the decimal to 27.5%. Which verification is required?

  1. Trust the generated summary because it contains the right general topic.
  2. Validate numeric values and units against the cited source spans before posting them to the system.
  3. Convert every percentage to a whole number to simplify downstream storage.
  4. Infer the rate from other customers’ contracts if the extraction appears unusual.

Explanation

Correct Answer

B:

Literal amounts and units must be grounded in the agreement to prevent consequential decimal shifts.

Incorrect Answers

A:

A topical match does not validate the precise percentage required for a fee.

C:

Rounding would corrupt a contractually significant fractional percentage.

D:

Unrelated agreements do not supersede the source document’s stated amount.

Question 681

A project review contains two contradictory status notes: engineering reports a completed feature while QA reports an unresolved blocking defect. Executives need a reliable summary. What should the model produce?

  1. Select the most positive status and omit the blocking QA note for readability.
  2. Average the two status claims into a fabricated intermediate delivery state.
  3. Describe the disagreement explicitly and attribute each status claim to its source.
  4. State that both reports are identical because they reference the same feature.

Explanation

Correct Answer

C:

Grounded attribution prevents the summary from silently collapsing conflicting evidence into one asserted fact.

Incorrect Answers

A:

Removing a material contradiction misrepresents the project state.

B:

Incompatible textual judgments cannot be resolved by arbitrary arithmetic.

D:

Referring to the same feature does not eliminate disagreement about its readiness.

Question 682

A field recording contains short spoken service notes separated by long periods of machinery silence. The speech pipeline spends unnecessary time trying to transcribe non-speech segments. Which preprocessing step is useful?

  1. Apply voice-activity detection to select likely speech regions before recognition.
  2. Treat every quiet section as a new speaker and transcribe it as a voice command.
  3. Increase the text generator completion limit to make silent segments recognizable.
  4. Delete all low-volume audio, including actual soft-spoken words, without evaluation.

Explanation

Correct Answer

A:

Identifying speech intervals avoids spending recognition effort on extended nonspeech portions.

Incorrect Answers

B:

A speaker label does not create spoken content where the recording contains silence.

C:

Generation token limits do not turn nonspeech audio into valid transcription.

D:

Blind amplitude removal can discard legitimate speech and degrade accuracy.

Question 683

A legal review team has a two-hour recorded interview. Speech recognition is accurate overall, but a few technical names have low-confidence word spans. Reviewers cannot listen to the entire file. What output best supports efficient correction?

  1. Return one confidence average for the full recording without identifying affected words.
  2. Replace low-confidence words with common alternatives before a reviewer hears the original.
  3. Remove timestamps because they are unrelated to transcript readability.
  4. Expose confidence-marked transcript spans with timestamps linked to the source audio.

Explanation

Correct Answer

D:

Reviewers can replay the uncertain words instead of manually inspecting every minute.

Incorrect Answers

A:

A single score does not locate the small number of problematic segments.

B:

Automatic substitution can corrupt names and hide genuine uncertainty.

C:

Without timing references reviewers cannot jump to the unclear audio.

Question 684

A speech assistant correctly transcribes general English but frequently misses a small fixed list of brand-specific device names. The team wants a lightweight first change before gathering a large custom training corpus. What should it evaluate?

  1. Immediately train a completely new acoustic model without testing a simpler contextual adaptation.
  2. Test supported phrase-list or contextual vocabulary hints with representative utterances.
  3. Replace every unfamiliar word with the most common English synonym without checking audio.
  4. Ignore the specialized names during evaluation and report general-language accuracy only.

Explanation

Correct Answer

B:

A constrained vocabulary hint can bias recognition toward known specialized terms without full custom training.

Incorrect Answers

A:

Full customization may be unnecessary for a short stable terminology list.

C:

Synonyms can change exact brand or model identifiers required by the application.

D:

Removing required vocabulary hides the real failure in production use.

Question 685

A voice assistant generates a long response incrementally. Users should hear the response quickly, but synthesizing each incoming token separately produces unnatural gaps and broken words. What output strategy should it use?

  1. Wait until the entire long response is complete before any audio is synthesized.
  2. Synthesize each raw token as a separate audio file and concatenate without prosody checks.
  3. Buffer coherent phrases or sentences and synthesize playable chunks in order.
  4. Send streamed text bytes directly to the speaker device as though they were audio samples.

Explanation

Correct Answer

C:

Chunking at language boundaries balances low initial latency with intelligible continuous speech.

Incorrect Answers

A:

Full buffering adds avoidable delay when the output supports incremental playback.

B:

Token-level synthesis fragments words and often damages natural pronunciation.

D:

Text encoding is not an audio waveform that can be played as speech.

Question 686

A support center records calls containing both Arabic and English phrases. Agents need a searchable record preserving each speaker’s actual wording, with optional translation for reviewers. Which processing output fits?

  1. Keep the original-language transcript with language tags and produce translations as separate fields.
  2. Store only a translated English paraphrase and discard the recognized source speech.
  3. Assign the entire conversation to one language based on the first spoken word.
  4. Generate new dialogue in both languages instead of transcribing the call.

Explanation

Correct Answer

A:

Preserving source words supports auditability while translations provide optional comprehension.

Incorrect Answers

B:

A paraphrase can lose exact customer wording and cannot be independently verified.

C:

Code-switching makes a single fixed language label potentially inaccurate.

D:

Fabricated conversation text is not evidence of what was said.

Question 687

A customer helpline must pronounce the product name “XQ-7” as specified by the manufacturer and slow down when reading troubleshooting steps. The chosen speech synthesizer supports markup controls. What should the integration supply?

  1. Change the speech-recognition language model for incoming calls and leave synthesis settings unchanged.
  2. Increase the audio file bit rate but supply no pronunciation or rate instructions.
  3. Translate the instruction into a different language before synthesis without preserving the product name.
  4. Use supported speech-synthesis markup for pronunciation and speaking-rate controls.

Explanation

Correct Answer

D:

Explicit pronunciation and prosody markup addresses the spoken output without altering factual text.

Incorrect Answers

A:

Recognition processes user audio and does not directly fix generated speech pronunciation.

B:

Higher encoding quality does not define how letters and numbers should be spoken.

C:

Translation can alter a protected technical identifier rather than controlling speech rendering.

Question 688

A compliance search app needs document IDs, section titles and short approved snippets for citation. Its index also contains confidential notes that the assistant does not need. How should query result fields be selected?

  1. Return every indexed field and ask the generator to omit confidential notes from the final answer.
  2. Project only required authorized fields from search results before constructing the model prompt.
  3. Convert secret note values to vector scores and treat scores as safe source citations.
  4. Increase top_k until the sensitive notes become a smaller fraction of the total response.

Explanation

Correct Answer

B:

A narrow result projection reduces unnecessary sensitive content leaving the retrieval boundary.

Incorrect Answers

A:

The model still receives data beyond its task requirements.

C:

A similarity score neither protects the original data path nor identifies valid source evidence.

D:

More retrieved content expands, rather than minimizes, possible exposure.

Question 689

A parts database must find model code AB-204-X exactly. The default text analyzer splits the hyphens and retrieves unrelated code fragments. Which index design should the team evaluate?

  1. Increase semantic embedding similarity until a related family member becomes the top match.
  2. Remove hyphens from every code without a validated canonical normalization rule.
  3. Use an appropriate exact-match field or analyzer strategy for the canonical part code.
  4. Raise the model response token limit while leaving identifier tokenization unchanged.

Explanation

Correct Answer

C:

Preserving identifier tokens enables reliable equality-style retrieval of the specific product code.

Incorrect Answers

A:

Semantic proximity is not proof of an exact inventory identifier match.

B:

Blind rewriting may merge identifiers that the source catalog treats as different.

D:

Generation capacity does not correct search analyzer behavior.

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.

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