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 54

Question 530

A multi-tenant search app retrieves the top five passages globally, then removes documents belonging to other tenants. Many relevant authorized passages are excluded because other tenants dominated the initial top five. What search-query change fixes this?

  1. Increase the generative model response token limit without changing the search query.
  2. Perform global top-k search first and ask the model to reconstruct filtered-out documents.
  3. Disable tenant authorization to increase the number of candidate passages.
  4. Apply the authorized tenant filter within retrieval before selecting and ranking top-k candidates.

Explanation

Correct Answer

D:

Pre-filtering retrieval to authorized documents allows ranking and top-k selection to operate over the correct tenant corpus.

Incorrect Answers

A:

Generation length cannot repair the search candidate set or the order in which authorization filters were applied.

B:

Filtering only after global top-k selection leaves too few authorized candidates; a model cannot recover missing evidence reliably.

C:

Removing access controls violates tenant confidentiality and is not an acceptable recall optimization.

Question 531

Hybrid search never returns the known supporting policy passage anywhere in its initial candidate set, although the clause is present in the index. Someone proposes semantic reranking. What must be fixed first?

  1. Increase the model completion budget without checking keyword or vector retrieval.
  2. Tune initial keyword/vector matching and filters so the missing passage enters the candidate set.
  3. Delete first-stage retrieval and expect semantic reranking to search the entire corpus directly.
  4. Apply more reranker weight to unrelated documents because the expected passage is absent.

Explanation

Correct Answer

B:

A semantic reranker can reorder only retrieved candidates; the candidate-stage recall failure must be repaired first.

Incorrect Answers

A:

Generation budget affects answer length, not whether the correct source passage is retrieved.

C:

A reranker requires an initial set of candidates and cannot replace retrieval over the source corpus.

D:

Reweighting unrelated candidates cannot introduce the missing passage into the ranked list.

Question 532

A procedure document contains numbered steps under section headings. Splitting at arbitrary sizes separates a step from its safety warning. What chunking policy fits?

  1. Index only document titles and omit the procedure text.
  2. Split every 100 characters without overlap or structural awareness.
  3. Use layout-aware sections that retain a step with its safety warning.
  4. Sort all headings alphabetically and discard step order.

Explanation

Correct Answer

C:

Preserving document structure improves retrieval of complete procedural context.

Incorrect Answers

A:

Titles cannot provide the needed safety conditions.

B:

Mechanical slicing can detach warnings from actions.

D:

Reordered procedures can be dangerous or nonsensical.

Question 533

A purchase-order table spans two pages, and naive text splitting assigns item totals to the wrong supplier row. What should ingestion preserve?

  1. Discard continuation pages after the first page of the table.
  2. Sort numbers by magnitude before attaching them to item descriptions.
  3. Flatten all cell values into one unordered paragraph.
  4. Use layout-aware chunking that preserves table row/column associations.

Explanation

Correct Answer

D:

Table relationships are essential for exact line-item interpretation.

Incorrect Answers

A:

Missing rows invalidate the purchase-order totals.

B:

Numeric ordering does not preserve original table semantics.

C:

Without structure values may attach to incorrect rows.

Question 534

A scanned legal PDF has two columns and footnotes. Fixed-length text slicing mixes unrelated clauses. Which preprocessing helps?

  1. Generate searchable chunks from document metadata without extracting the PDF page content.
  2. Apply OCR and section-aware chunking for scanned contracts.
  3. Use raw image byte offsets as if they were text clause boundaries.
  4. Alphabetize OCR output to avoid repeated words.

Explanation

Correct Answer

B:

Reading order and spatial relationships should survive extraction for faithful grounding.

Incorrect Answers

A:

Titles and file attributes omit the legal clauses the assistant must retrieve.

C:

Compressed bytes do not reflect readable legal sections.

D:

Sorting destroys sentences and column sequence.

Question 535

An Azure AI Search indexer must create smaller text chunks from long ingested documents for vectorization. Which indexing enrichment is relevant?

  1. Use semantic-ranker configuration as the only document splitter.
  2. Use a text-splitting skill in a search indexer pipeline.
  3. Increase chat model temperature to split input files automatically.
  4. Store complete multi-hundred-page documents as one opaque vector without evaluating limits.

Explanation

Correct Answer

B:

Text splitting prepares manageable chunks for indexing and embedding.

Incorrect Answers

A:

Reranking happens at query time, not as the initial chunking operation.

C:

Sampling controls do not define indexer text segments.

D:

Oversized unsegmented content can harm indexing and relevance.

Question 536

A search pipeline ingests blobs and should automatically turn document chunks into vectors using a supported Azure integration. Which capability should be configured?

  1. Only set model output-token limits and assume search vectors appear.
  2. Use integrated vectorization with compatible embedding skills, field mappings and index schema.
  3. Store the blob container name as each document's embedding.
  4. Embed source text with a different incompatible model and insert the result into the existing vector field.

Explanation

Correct Answer

B:

An integrated pipeline can generate indexed embeddings during ingestion.

Incorrect Answers

A:

Generation parameters do not configure indexing vectorization.

C:

A metadata string is not a numerical semantic vector.

D:

Mismatched embeddings do not preserve the representation contract required between indexed vectors and queries.

Question 537

A data science team already computed approved embeddings outside Azure AI Search. It needs to import them into an index. What should ingestion do?

  1. Store the embedding file's path string as a vector value.
  2. Use externally generated vectors when ingesting precomputed embeddings.
  3. Re-embed the indexed documents with an unrelated model while keeping query vectors in their old embedding space.
  4. Use a generative response endpoint as the index storage route.

Explanation

Correct Answer

B:

External vectorization is usable when schema and representation consistency are enforced.

Incorrect Answers

A:

A path is not the numeric array required for vector search.

C:

Matching dimensions alone are not sufficient: query and document vectors must use compatible semantics.

D:

Inference APIs do not replace index document upload.

Question 538

A query vector contains 3072 dimensions but the search index field expects 1536. What must be corrected?

  1. Increase semantic ranker top-k instead of fixing the vector field mismatch.
  2. Truncate each 3072-dimensional query vector to 1536 elements without regenerating compatible embeddings.
  3. Truncate or pad arbitrary numbers without checking model compatibility.
  4. Use matching dimensionality and model-compatible query/index embeddings.

Explanation

Correct Answer

D:

Vector query and field dimensions must match for similarity operations.

Incorrect Answers

A:

Reranking cannot operate correctly on invalid vector input.

B:

Arbitrary truncation changes vector meaning and does not establish a valid index/query embedding representation.

C:

Ad hoc numeric changes can invalidate semantic meaning.

Question 539

A knowledge-base app must search by words, compare vectors and show the exact cited passage. What index schema design supports all three?

  1. Store only full-text fields and expect vector queries without embeddings.
  2. Create searchable text, vector fields and retrievable source metadata.
  3. Use record identifiers as vector content without computing semantic embeddings.
  4. Store only vectors and discard every original content field.

Explanation

Correct Answer

B:

Each query and citation function requires an appropriate representation in the index.

Incorrect Answers

A:

Text fields alone do not satisfy a vector search contract.

C:

Identifiers do not capture document meaning for nearest-neighbor retrieval.

D:

Vectors cannot supply readable evidence passages alone.

Question 540

Search returns a correct paragraph, but the assistant has no link for users to open the original policy. Which field should indexing preserve?

  1. Choose a retrievable document URL field for citation generation.
  2. Delete document IDs after chunk vectorization.
  3. Compose citation URLs from predictable intranet patterns without resolving actual records.
  4. Store only a cosine similarity score and treat it as a hyperlink.

Explanation

Correct Answer

A:

Citation links require resolvable source provenance independent of generated prose.

Incorrect Answers

B:

Removing source keys breaks citation and lifecycle tracking.

C:

A plausible-looking URL is not authoritative source provenance.

D:

A score is not a document locator.

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