cert
cert-1
cert-2

Pass Microsoft AI-103 Exam in First Attempt Guaranteed!

Get 100% Latest Exam Questions, Accurate & Verified Answers to Pass the Actual Exam!
30 Days Free Updates, Instant Download!

cert-5
cert-6
AI-103 Exam - Verified By Experts
AI-103 Premium File

AI-103 Premium File

$89.99
$98.99
  • Premium File 187 Questions & Answers. Last Update: Nov 15, 2026

Whats Included:

  • Latest Questions
  • 100% Accurate Answers
  • Fast Exam Updates
 
$98.99
$89.99
accept 397 downloads in the last 7 days
block-screenshots
Microsoft AI-103 Practice Test 7

Question 60

Meeting transcripts mix customer and employee statements. What metadata should be indexed?

  1. Store one meeting-level speaker label for an entire recording with multiple participants
  2. Store speaker or channel attribution with transcript spans
  3. Store all transcript turns in one unlabelled text field and infer each speaker solely from nearby punctuation at search time.
  4. Transcribe audio and index timestamped text segments

Explanation

Correct Answer

B:

Speaker identity is essential when retrieving who said a particular phrase. In this scenario the team must separate speakers in indexed audio. A feature used to navigate from text hits to audio evidence does not substitute for it.

Incorrect Answers

A:

One global speaker label erases attribution across turns. Transcript segments need speaker or channel metadata aligned with each utterance.

C:

Merged text loses reliable attribution when callers interrupt or switch turns. Index the speaker or channel identifier with each transcript span and timestamp.

D:

Transcripts make spoken content text-searchable while timestamps preserve links to the recording. This helps when a team needs to create searchable audio transcripts; it does not establish how to separate speakers in indexed audio.

Question 61

A support RAG index contains 400-page manuals stored as a single vector each, causing irrelevant context. What ingestion design helps?

  1. Embed each manual as one large vector and retrieve the entire book.
  2. Index only the manual titles and leave the full body outside retrieval.
  3. Increase the language model's output-token limit without changing ingestion.
  4. Chunk documents into semantically coherent passages with source and section metadata.

Explanation

Correct Answer

D:

Passage-level retrieval can surface the relevant section while preserving references to its document and version.

Incorrect Answers

A:

Whole-document vectors are too coarse for locating a precise answer inside hundreds of pages.

B:

Titles cannot provide the specific technical passage needed to ground the response.

C:

An output budget does not fix low-granularity retrieval and irrelevant source context.

Question 62

An application requires audio transcription and natural-sounding spoken responses. Which service capability should be integrated?

  1. Use Azure Speech capabilities for speech recognition and synthesis
  2. Call a CRM lookup function and assume its text response performs speech recognition and spoken output.
  3. Index text transcripts in a search service and assume that automatically produces spoken audio responses.
  4. Use a document-field extraction analyzer to produce speech audio directly without speech recognition or synthesis APIs.

Explanation

Correct Answer

A:

Speech-to-text and text-to-speech provide the needed audio interfaces for an agentic user experience. This resolves the specific need to convert speech to text and text back to audio, while a measure that helps connect an agent to a live protected business API addresses a different requirement.

Incorrect Answers

B:

A CRM tool returns business data, not audio-to-text transcription or synthesized speech. The request needs supported speech recognition and text-to-speech processing.

C:

Search indexing cannot perform speech recognition or synthesis. A voice interaction requires audio transcription plus speech generation APIs with appropriate language and format settings.

D:

Document extraction identifies structured information from content; it does not implement both speech-to-text and text-to-speech. Integrate Azure Speech recognition and synthesis capabilities for the conversation.

Question 63

An editor wants to change a model's generated image globally from midday to sunset without selecting a specific object. What input approach fits?

  1. Mask the outer border for canvas extension while leaving the central scene's lighting unchanged, even though sunset color is required throughout.
  2. Apply inpainting to a small foreground-object mask while leaving the entire sky and scene illumination unedited.
  3. Limit the change to a small object mask while leaving the sky and scene lighting untouched.
  4. Use a prompt-driven global image-to-image editing operation

Explanation

Correct Answer

D:

Global edits affect the full scene's look, unlike a localized mask. This resolves the specific need to adjust overall image lighting or style, while a measure that helps extend the image beyond its existing frame addresses a different requirement.

Incorrect Answers

A:

Outpainting addresses missing canvas edges, not an intentional global lighting change. A supported image-to-image edit with a sunset prompt is the relevant operation.

B:

A local foreground mask cannot reliably transform the whole image from midday lighting to sunset. The required operation is a globally conditioned image-to-image edit, not isolated object replacement.

C:

A tiny object mask does not change the global lighting required for a midday-to-sunset transformation. The image-wide prompt or edit conditioning must address the whole scene.

Question 64

An application must detect private passport numbers inside chat before summarization. Where should the protection run?

  1. Redact passport identifiers only from the final visible summary.
  2. Treat every number as harmless if the summarizer promises not to repeat it.
  3. Perform sensitive-information detection and redaction before text enters the summary model context.
  4. Log complete raw chats for later review before applying privacy rules.

Explanation

Correct Answer

C:

Preprocessing reduces exposure of regulated identifiers in prompts and downstream logs rather than depending on the generated summary to omit them.

Incorrect Answers

A:

The sensitive value would already have entered model prompts and potentially internal telemetry before output filtering.

B:

Instructions do not prevent exposure of sensitive numbers inside the model context.

D:

Storing unredacted passport data increases exposure and does not satisfy preprocessing privacy needs.

Question 65

A contract table has invoice amounts under several headers, but the extraction pipeline loses the relationship between cells. Which representation is required?

  1. Flatten the entire table into a single paragraph without cell positions.
  2. Return just the list of amounts sorted from smallest to largest.
  3. Preserve table row, column, header and cell relationships in structured extraction output.
  4. Preserve only column titles without their corresponding rows and cells.

Explanation

Correct Answer

C:

Raw flattened text can destroy tabular meaning; downstream questions require the correct field-to-value associations.

Incorrect Answers

A:

Flattening discards which monetary amount belongs under which header or row.

B:

Sorting amounts by value discards the invoice row, column header and original cell relationships. The extraction result must retain table geometry and the association between an amount and the corresponding invoice item.

D:

Headers alone cannot identify the correct amount for a particular invoice line.

Question 66

A retriever returns ten passages, but the app feeds them to the model without indicating boundaries or sources. What should change?

  1. Insert retrieved paragraphs as undelimited plain text and attach a single generic citation to the entire prompt after generation.
  2. Wrap retrieved snippets with explicit source identifiers and delimiters in the grounded prompt
  3. Concatenate passages without provenance and grant the first retrieved document authority to command agent tool calls.
  4. Enable hybrid ranking but remove source boundaries and document identifiers from every supplied passage.

Explanation

Correct Answer

B:

Source separation lets generation and citation checks distinguish evidence from instructions and unrelated passages. In this scenario the team must structure retrieved context with citations. A feature used to test whether citations support the answer does not substitute for it.

Incorrect Answers

A:

Without per-passage boundaries and source IDs, the model cannot reliably associate facts with their origins. Delimited evidence with explicit provenance is needed during prompt assembly.

C:

Missing document boundaries and source references weakens attribution, while promoting source text to instructions creates prompt-injection risk. Preserve passage identity and isolate untrusted text.

D:

Ranking improves which passages are found, not their provenance within the model context. Pass bounded excerpts with unambiguous source IDs and citations to support grounded answers.

Question 67

A developer wants one project context to select and invoke deployed generative models while managing shared credentials and settings. What resource should be configured?

  1. Create an unversioned local chat transcript to substitute for an authenticated Foundry project and its deployed model connection.
  2. Configure a Microsoft Foundry project with an appropriate model deployment
  3. Use a document analysis job as though it automatically creates a Foundry project and model inference deployment.
  4. Create an Azure AI Search index and assume it establishes a Foundry project, model deployment, and inference route.

Explanation

Correct Answer

B:

The Foundry project supplies a coherent development context for consuming and governing deployed models. The practical requirement is to manage model deployments and project-scoped connections; a solution built to maintain agent continuity with grounded knowledge would not meet that same need.

Incorrect Answers

A:

Conversation state does not configure the serving model or the project's authorized resources. The application needs the real Foundry project context and a corresponding deployment connection.

C:

Extracting fields from a document is unrelated to establishing a project context and deployed generative model. Configure the actual Foundry project, connection and model deployment required by the SDK.

D:

A search index can supply grounding context but is not a Foundry project or deployed model. Configure the appropriate project and model deployment, then call the documented inference endpoint through a compatible client.

Question 68

A knowledge assistant cites a document that is unrelated to the actual answer. What evaluation should be used?

  1. Accept the assistant's citation because a retrieved document title matches, without checking whether its passage supports the factual answer.
  2. Tune hybrid retrieval scores without checking whether the generated claim is supported by the cited passage.
  3. Check answer-to-citation support on a labeled grounding test set
  4. Obey the retrieved document's instruction to cite its title even when the document does not support the claim.

Explanation

Correct Answer

C:

Answer evidence needs claim-level support; relevant retrieval alone does not validate fabricated citations. This resolves the specific need to test whether citations support the answer, while a measure that helps structure retrieved context with citations addresses a different requirement.

Incorrect Answers

A:

A citation's existence does not demonstrate entailment. Evaluation must compare the answer's material claims against the retrieved and cited source passages.

B:

A high retrieval score does not prove the answer's cited evidence entails its claims. Evaluate citation relevance and grounded answer faithfulness against the actual sources.

D:

An irrelevant document cannot become valid evidence because it requests a citation. Compare each generated claim with the authorized passage and suppress unsupported attribution.

Question 69

An agent must call a protected inventory API to answer stock questions. Which integration capability does it need?

  1. Provide a plain text description of inventory API fields but do not register an authenticated executable tool or function schema.
  2. Use indexed product descriptions as the sole authority for live inventory quantities without invoking the protected stock API.
  3. Upload a screenshot of inventory into a document parser instead of authorizing a callable live stock API tool.
  4. Provide a defined callable API or function tool with authorized access

Explanation

Correct Answer

D:

A tool contract allows the agent controller to obtain current inventory from a system of record rather than inventing values. Here the relevant capability must let the team connect an agent to a live protected business API. An alternative targeting convert speech to text and text back to audio solves another problem.

Incorrect Answers

A:

Merely mentioning an API in a prompt does not authorize or execute its calls. The agent requires a defined tool contract and credentials scoped to permitted stock lookups.

B:

A cached knowledge index cannot guarantee current stock or enforce the inventory system's authorization contract. The agent needs an approved callable tool/API to request and validate live quantities at runtime.

C:

Document parsing can recover fields from uploaded content, but it cannot guarantee current inventory data or make authorized changes to the protected stock system. Integrate a scoped API tool.

Question 70

An application already uses hybrid keyword-and-vector retrieval, but exact error identifiers still rank below generic symptom matches. What search-ranking adjustment should the team test?

  1. Return every retrieved document without ranking or truncation to avoid missing information.
  2. Increase vector-similarity weight for all generic symptom documents and remove the field-specific boost for exact error identifiers.
  3. Wrap retrieved snippets with explicit source identifiers and delimiters in the grounded prompt
  4. Boost an exact-code field in the lexical branch and evaluate the blended ranking against both identifier and symptom queries.

Explanation

Correct Answer

D:

The team already has hybrid retrieval, so selecting hybrid search again would not fix the ranking. Increase the influence of exact-code matches in the lexical field or scoring profile, then measure precision for exact identifiers and recall for natural-language symptoms. The existing semantic branch should still retrieve paraphrases. Lexical field boosting can prioritize precise error-code matches while the existing vector branch continues retrieving semantically similar descriptions.

Incorrect Answers

A:

Unbounded evidence increases noise and context cost; it does not improve exact-code relevance at the top of an already hybrid ranked result set.

B:

Semantic similarity can bury unique code tokens beneath broad symptom text. Exact-code fields need lexical weighting while the blended ranking is evaluated on both query types.

C:

Source separation lets generation and citation checks distinguish evidence from instructions and unrelated passages. This is a separate capability and does not directly tune exact-identifier influence within an existing hybrid index.

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