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 10

Question 90

A finance team compares token expenses for prompts containing 20,000 tokens versus responses containing 500 tokens. The price schedule differentiates input and output. What calculation should it use?

  1. Use the response character count as the entire bill without measuring billed token units.
  2. Charge only on the number of HTTP requests, ignoring token usage.
  3. Price input and output tokens separately and sum the charges.
  4. Multiply all tokens by the output rate without distinguishing categories.

Explanation

Correct Answer

C:

Prompt and completion tokens can carry different unit pricing; both terms contribute to total cost.

Incorrect Answers

A:

Characters do not map directly to billable input/output token consumption.

B:

Request count alone omits variable prompt and completion volume.

D:

Applying one rate to both categories can misstate costs under differential pricing.

Question 91

An internal assistant retrieves 20 nearly identical policy passages per answer and prompt expense doubles. Evidence audits show five passages are sufficient. Which change is appropriate?

  1. Limit redundant retrieved chunks according to relevance metrics.
  2. Increase retrieved passages to 50 to make the model less uncertain.
  3. Remove all citations but keep the same 20 passages in prompts.
  4. Increase maximum output tokens to compensate for expensive input.

Explanation

Correct Answer

A:

Fewer high-value passages can reduce input tokens while preserving answer support.

Incorrect Answers

B:

More context further raises token costs and may obscure relevant evidence.

C:

Citation formatting has little effect on the excess retrieved-token volume.

D:

A larger output allowance adds potential cost and does not reduce prompt consumption.

Question 92

A ticket triage service routes simple labels through a costly large model even though a smaller model meets acceptance accuracy on a representative test set. What optimization fits?

  1. Disable evaluation to prevent seeing errors introduced by a cheap model.
  2. Change all labels to a single default class to eliminate model calls.
  3. Increase model context for every short ticket instead of reconsidering its size.
  4. Route low-risk classification tasks to the evaluated smaller model and monitor label quality.

Explanation

Correct Answer

D:

Risk and task evidence support lower-cost inference where quality remains sufficient.

Incorrect Answers

A:

Without monitoring the stated quality requirement cannot be maintained.

B:

A constant output sacrifices required classification accuracy.

C:

Longer prompts may increase cost without solving the model-selection inefficiency.

Question 93

A support agent's p95 response latency increased after adding a search tool. A single total-duration metric cannot show which component regressed. What instrumentation is needed?

  1. Remove all dependency trace identifiers to avoid a noisy dashboard.
  2. Record only the time users log in and assume it explains inference delays.
  3. Increase generation variability and use answer length as the only latency performance measurement.
  4. Track model-call latency by retrieval, tool, and generation stages.

Explanation

Correct Answer

D:

Per-component spans localize latency to search, tool or inference operations rather than guess from a total.

Incorrect Answers

A:

Removing correlation makes the root cause less observable.

B:

Login timing does not identify which step of the agent request slowed down.

C:

Sampling and response length do not isolate time spent in retrieval, tools or generation.

Question 94

After a policy-corpus update, chatbot answers become less relevant although HTTP success remains high. Which quality signal will expose the regression?

  1. Measure CSS render time instead of model or retrieval quality.
  2. Compare answer relevance on a labeled baseline across document versions.
  3. Count server restarts and infer changes in answer relevance from the count.
  4. Use HTTP status 200 as sole proof that answers remain correct.

Explanation

Correct Answer

B:

A task-level benchmark can detect semantic-quality drift despite healthy transport metrics.

Incorrect Answers

A:

UI paint metrics cannot identify semantic regressions introduced by source changes.

C:

Process restarts are not a direct measure of semantic answer quality.

D:

A successful request says nothing about whether its content meets user intent.

Question 95

A RAG assistant writes a policy conclusion not present in any retrieved passage. Which evaluator most directly flags the unsupported claim?

  1. Evaluate groundedness against the retrieved source passages.
  2. Measure request rate and treat a low rate as evidence of correctness.
  3. Check only whether the answer uses a friendly tone.
  4. Calculate keyword popularity of the generated response without comparing citations.

Explanation

Correct Answer

A:

Groundedness assesses whether an answer's factual claims have supporting evidence in the retrieved material.

Incorrect Answers

B:

Traffic volume is unrelated to whether the answer is supported.

C:

Tone does not establish factual support from the provided documents.

D:

Common words can still occur in unsupported or fabricated claims.

Question 96

A user receives an incorrect benefits answer. Logs show the correct policy clause was never retrieved, so the model generated from irrelevant passages. Which fault should be investigated first?

  1. Disable search entirely and require the model to rely on its general training.
  2. Change model sampling temperature without investigating whether the cited document was retrieved correctly.
  3. Increase model temperature and assume absent documents will become visible.
  4. Investigate retrieval coverage and ranking before assigning the incident solely to hallucination in generation.

Explanation

Correct Answer

D:

Missing source evidence indicates a retrieval failure that downstream language generation cannot reliably overcome.

Incorrect Answers

A:

The correct policy resides in the private corpus and must be retrieved.

B:

Sampling controls cannot diagnose missing or irrelevant retrieval evidence.

C:

Sampling changes do not retrieve the missing policy clause.

Question 97

New knowledge-base articles are committed hourly, but the index has not refreshed since yesterday. Users ask why recent fixes are absent. What should be monitored?

  1. Monitor only how quickly the chatbot animates its typing indicator.
  2. Turn off source-version tracking so the index appears consistent.
  3. Increase output-token allowance and expect missing articles to appear.
  4. Monitor search index freshness after upstream document updates.

Explanation

Correct Answer

D:

Freshness metrics reveal when available upstream changes have not reached the retrievable index.

Incorrect Answers

A:

UI animation timing does not reflect whether new documents are indexed.

B:

Hiding revisions obscures stale content and undermines correctness.

C:

Output length has no effect on the search ingestion lifecycle.

Question 98

After a vendor changes PDF metadata fields, an Azure AI Search indexer stops processing new files. Existing queries still work. What should troubleshooting inspect?

  1. Assume successful queries mean the indexer is also healthy.
  2. Raise chat model temperature as the first response to indexer failures.
  3. Examine indexer execution errors and source-to-index field mappings for schema incompatibility.
  4. Rebuild the UI without reviewing the failing indexer logs.

Explanation

Correct Answer

C:

A source schema change can break ingestion independently of serving previously indexed documents.

Incorrect Answers

A:

A query can succeed against stale data while fresh indexing repeatedly fails.

B:

Generation settings cannot correct ingestion field-mapping exceptions.

D:

A frontend refresh does not explain a backend document-processing error.

Question 99

A pipeline processed 9,600 items, but the source folder holds 10,000 policy documents. The deployment dashboard is green. Which measure identifies the gap?

  1. Measure document-ingestion completeness against the source inventory.
  2. Measure average output tokens and ignore missing source identifiers.
  3. Treat a green web-server heartbeat as proof every source file entered the index.
  4. Exclude failed documents from the denominator without recording why they failed.

Explanation

Correct Answer

A:

Ingestion completeness requires matching expected inputs to successfully processed documents, not just service availability.

Incorrect Answers

B:

Model tokens do not report which documents were ingested.

C:

Web availability cannot prove data ingestion completeness.

D:

Silently shrinking the expected corpus conceals genuine coverage gaps.

Question 100

After an embedding upgrade, queries fail because stored vectors have 1536 dimensions but new request vectors use another size. Which check matters?

  1. Interpret missing vector matches as evidence that all documents are irrelevant.
  2. Validate embedding model and vector field dimensional compatibility, rebuilding vectors or schema if necessary.
  3. Increase the chat response token limit until dimensions agree.
  4. Change document titles to shorter strings and retry the same vector query.

Explanation

Correct Answer

B:

Vector similarity operations require compatible dimensions between indexed representations and queries.

Incorrect Answers

A:

A schema mismatch is a technical failure, not a relevance judgment.

C:

Output-token settings cannot change vector length compatibility.

D:

Metadata text length does not resolve numeric vector dimension mismatches.

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