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 23

Question 220

A JSON extraction prompt produces inconsistent field names across runs. Developers want stable structure without creating a custom model. What prompt technique may help?

  1. Use task-specific few-shot examples paired with a strict output schema and validation.
  2. Allow the model to invent new fields freely for each record.
  3. Set the maximum response length to one token and omit examples.
  4. Use broad generic prompt examples instead of examples that demonstrate the target JSON structure.

Explanation

Correct Answer

A:

Relevant demonstrations clarify desired extraction patterns while schemas enforce the receiving contract.

Incorrect Answers

B:

Unbounded keys do not meet downstream parser requirements.

C:

A tiny output budget cannot express the required structured record.

D:

Unrelated demonstrations do not teach the required field names and typing contract.

Question 221

A procurement extractor uses static examples of valid invoices, followed by live untrusted invoice text. A supplier inserted instructions telling the model to change its schema. How should inputs be separated?

  1. Separate trusted few-shot instructions from untrusted live input.
  2. Run any tool command embedded in the invoice before extracting its fields.
  3. Place supplier text into the system instruction as if it were an approved example.
  4. Omit application-side field checks because few-shot examples guarantee compliance.

Explanation

Correct Answer

A:

Role and data boundaries prevent external text from redefining extraction rules.

Incorrect Answers

B:

Source text is evidence to parse, not authorization for executable actions.

C:

An external document must not gain privileged instruction authority.

D:

Demonstrations improve consistency but do not prove every output valid.

Question 222

A team modifies a prompt template and sees fewer customer complaints but no measured before/after quality data. What release evidence should it keep?

  1. Use output length as the sole measure of better answers.
  2. Track prompt-template revisions with test result comparisons.
  3. Change model and retrieval configuration simultaneously without recording either change.
  4. Overwrite the only prompt copy and discard prior evaluation outcomes.

Explanation

Correct Answer

B:

Versioned tests identify regressions attributable to prompt changes rather than intuition alone.

Incorrect Answers

A:

Longer answers need not be more correct or useful.

C:

Untracked simultaneous changes obscure cause and effect.

D:

Without history reviewers cannot explain changes or roll back reliably.

Question 223

A developer's few-shot examples include a real API credential to demonstrate the expected JSON format. How should examples be sanitized?

  1. Put additional admin keys into examples so the model knows fallback accounts.
  2. Use synthetic nonsecret values and approved schemas; remove credentials from prompt examples and stored traces.
  3. Put an encoded credential into a few-shot prompt and rely on output redaction to conceal it.
  4. Keep the live credential because examples are never stored by services.

Explanation

Correct Answer

B:

Examples should teach structure without disclosing reusable secrets to the model or logs.

Incorrect Answers

A:

Additional credentials increase the exposure risk without improving the schema.

C:

The credential is still exposed to the prompt-processing path even if the displayed answer is filtered.

D:

Prompt and diagnostic systems may persist or expose submitted content.

Question 224

A user reports a wrong answer; support needs to reconstruct the associated search and model requests. Separate systems have unrelated logs. What identifier should the app supply?

  1. Infer missing request IDs from generated answer text after the incident.
  2. Use the same static ID for every user request to simplify filtering.
  3. Store an application request ID for correlated model and search traces.
  4. Record only the model family name for all failures.

Explanation

Correct Answer

C:

Shared IDs make one interaction traceable across service boundaries and error events.

Incorrect Answers

A:

A model response cannot recreate authoritative trace identifiers.

B:

A global ID cannot distinguish separate interactions accurately.

D:

Model family does not uniquely identify the problematic request.

Question 225

A RAG assistant's retrieved passages exceed the model context even though each individual chunk is small. Which constraint should retrieval enforce?

  1. Set output max tokens higher while leaving oversized input unchanged.
  2. Drop source IDs but send every long passage verbatim.
  3. Budget retrieved-context tokens before the generation step.
  4. Increase top-k without measuring passage lengths.

Explanation

Correct Answer

C:

A retrieval token budget keeps combined evidence within model context limits.

Incorrect Answers

A:

Completion limits do not remove an overfull prompt context.

B:

Removing identifiers barely addresses the main evidence volume and loses provenance.

D:

More chunks may worsen the input overflow.

Question 226

An internal search index has a brief outage. The assistant supports a separately validated backup corpus containing the latest approved policies. What response path is appropriate?

  1. Silently return last year's policy without indicating it is stale.
  2. Open an unrestricted external search site for confidential policy content.
  3. Have the model fabricate source passages while the index is offline.
  4. Use an authorized fallback retrieval source with freshness checks, or return an explicit unavailable result.

Explanation

Correct Answer

D:

A verified fallback preserves evidence quality without pretending missing data was retrieved.

Incorrect Answers

A:

Stale unqualified results can mislead users on current entitlements.

B:

Unauthorized sources cannot replace the protected internal corpus.

C:

Invented evidence can yield false policy answers.

Question 227

An application must answer questions and retrieve semantically similar contract clauses. The team considers deploying only one chat model for both tasks. Which deployment split is sound?

  1. Send document text to the chat model and store its deployment name as a vector.
  2. Use embeddings as the final legal-language response without synthesis.
  3. Choose separate model deployments for inference and embeddings.
  4. Use the text-generation deployment alone and estimate similarity from its response wording.

Explanation

Correct Answer

C:

Embedding and generation models serve different API capabilities and should be configured accordingly.

Incorrect Answers

A:

A model identifier is not a semantic embedding representation.

B:

Vectors do not provide a user-readable answer.

D:

A generation endpoint is not an appropriate substitute for consistent document/query embeddings and vector retrieval.

Question 228

A business agent must call a structured inventory tool, but the chosen model deployment does not support the required tool-calling interface. What should be verified?

  1. Parse arbitrary prose and execute it as shell commands regardless of authorization.
  2. Set temperature to zero to add unsupported tool-calling features.
  3. Increase search index size so the model gains tool-call syntax.
  4. Check the deployed model supports structured tool calling.

Explanation

Correct Answer

D:

Tool capability is model- and version-dependent and must be established before workflow design.

Incorrect Answers

A:

Free-form execution is not a safe substitute for a validated tool interface.

B:

Sampling changes do not alter the model API's supported capabilities.

C:

Search capacity does not add inference API features.

Question 229

A compliance ingestion system accepts records with typed `rule_id` and `effective_date`; an LLM sometimes replies with commentary instead. What output design fits?

  1. Ask for maximum creativity with no field restrictions.
  2. Require schema-constrained JSON and validate typed compliance fields.
  3. Store free-form essays as date fields and let the database infer values.
  4. Use model fluency as the only validation criterion.

Explanation

Correct Answer

B:

A machine-readable contract can reliably feed a typed compliance pipeline.

Incorrect Answers

A:

Unconstrained output increases parser variability and missing required values.

C:

Narratives are not guaranteed valid typed dates or IDs.

D:

Well-written text can violate the receiving system's schema.

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.

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