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 1

Question 1

A Python application has a Foundry deployment ID and needs to generate a response to a user prompt. Which implementation step actually invokes it?

  1. Define a strict output JSON schema and then call only a project-management operation, assuming the schema will invoke the deployed model.
  2. Call the deployed model through the supported SDK or inference endpoint using its deployment identifier
  3. Increase the completion-token limit while continuing to send all requests exclusively to a project administration endpoint instead of model inference.
  4. Use bounded concurrency and retry-after-aware backoff for model requests

Explanation

Correct Answer

B:

A deployed model must be addressed through its serving API; a project name alone does not produce an inference. Here the relevant capability must let the team invoke the configured model endpoint. An alternative targeting enforce a structured response shape solves another problem.

Incorrect Answers

A:

A structured-output schema controls the response shape after inference. It does not send a prompt to the model-serving endpoint; use the supported SDK or inference route with the correct deployment identifier.

C:

Token limits change generation parameters only after a model has been invoked. They cannot make a project management endpoint process an inference request.

D:

Client-side admission and backoff prevent synchronized floods during provider throttling. Here the needed outcome is to invoke the configured model endpoint; an option designed to manage model request concurrency and throttling would not directly provide it.

Question 2

A branch kiosk has already passed accuracy tests for its fixed-label classifier, but must run during network outages with limited GPU memory. Which deployment model choice meets that requirement?

  1. Keep the kiosk dependent on a live cloud pricing-retrieval endpoint even though the classifier must work without a network connection.
  2. Deploy a compact language model suitable for local inference
  3. Deploy a very large cloud-only image-recognition model and depend on uninterrupted internet access for fixed-label text classification.
  4. Test a smaller model against the fixed-label accuracy target before selecting a large model

Explanation

Correct Answer

B:

A small model can meet bounded classification needs with a smaller memory and latency footprint, while local execution avoids cloud connectivity dependencies. Here the relevant capability must let the team run short classifications with limited hardware and no network. An alternative targeting separate language generation from live authoritative data solves another problem.

Incorrect Answers

A:

Network-dependent retrieval cannot serve an outage-tolerant kiosk. The fixed-label classifier should run locally on a model validated for accuracy within the available memory budget.

C:

The requirement is offline text classification on limited memory, not remote image recognition. A compact locally runnable model is the appropriate starting point if its accuracy tests pass.

D:

Structured narrow tasks may not require a general large model; accuracy, latency and measured spend should determine the smallest adequate choice. Here the needed outcome is to run short classifications with limited hardware and no network; an option designed to pick sufficient capability for a bounded output taxonomy would not directly provide it.

Question 3

A marketing team wants a new illustration generated solely from a written scene description. What operation should the application invoke?

  1. Handle a normal text-to-image request as a policy refusal without invoking the image generation capability.
  2. Generate an image from a supported text prompt using an image-capable model
  3. Submit the scene description to an image-captioning endpoint that can analyze existing photographs but cannot synthesize a new illustration.
  4. Store an output hash and prompt ID without making a text-to-image generation request.

Explanation

Correct Answer

B:

Text-to-image synthesis creates novel visual content from a natural-language description. Here the relevant capability must let the team create a synthetic image from prose instructions. An alternative targeting handle disallowed generated-image requests safely solves another problem.

Incorrect Answers

A:

A valid scene prompt requires an image-capable synthesis operation. A refusal-handling branch is useful only when the actual request fails content policy.

C:

Image captioning describes supplied pixels; the marketing request contains only a written scene and requires new imagery, so a text-to-image generation endpoint is needed.

D:

Provenance is useful after an asset exists; it does not synthesize an image from text. Invoke a supported image-generation operation with the description as its prompt.

Question 4

An insurer needs names of organizations and policy numbers mentioned in adjuster notes. Which text-analysis output is most relevant?

  1. Run a sentiment classifier over the entire adjuster note and treat its positive or negative score as the extracted policy number.
  2. Extract entities together with their contextual roles or relations
  3. Extract named entities and domain-defined identifier fields
  4. Supply local sentence context and domain-specific disambiguation criteria

Explanation

Correct Answer

C:

Entity extraction identifies spans that represent people, organizations and specific domain values. The decisive difference is whether the design can identify entity spans and policy identifiers; the competing function serves to discover recurring themes without fixed labels.

Incorrect Answers

A:

Sentiment scores describe tone, not named organizations or policy identifiers. They cannot return the exact entity spans or domain-defined ID values the claims workflow requires.

B:

Entity names alone do not describe who did what to whom in a sentence. Here the needed outcome is to identify entity spans and policy identifiers; an option designed to capture relationships between named entities would not directly provide it.

D:

Ambiguous terms require surrounding evidence and taxonomy rules to resolve the intended sense. That is relevant to disambiguate entity meanings using context, but the present scenario requires an ability to identify entity spans and policy identifiers.

Question 5

A collection of PDF policies has been ingested as raw files, but its text has not yet been extracted into any searchable representation. What initial transformation is needed before chunking and indexing?

  1. Assign a unique random storage filename to each PDF but never extract its text or preserve a source document identifier.
  2. Track source deletions and remove or mark deleted index entries
  3. Extract and normalize document text with source identifiers before indexing
  4. Split documents into semantically coherent chunks with stable source offsets

Explanation

Correct Answer

C:

Normalized text and provenance enable later retrieval and citation of relevant policy passages. Here the relevant capability must let the team prepare document text and identity for retrieval. An alternative targeting deduplicate recurrent document ingestion solves another problem.

Incorrect Answers

A:

Renaming stored files does not create searchable policy passages. Extract and normalize the actual content, associate it with stable provenance, then chunk and index it.

B:

Ingestion must reconcile removed source content so responses do not cite obsolete files. The function serves to synchronize indexed deletions, whereas the actual requirement is to prepare document text and identity for retrieval.

D:

Well-designed chunks improve retrieval specificity while retaining citation lineage. Here the needed outcome is to prepare document text and identity for retrieval; an option designed to index targeted passages rather than whole manuals would not directly provide it.

Question 6

A knowledge-backed assistant quotes an obsolete employee benefits policy after a newer revision was indexed. Which retrieval metadata must be enforced?

  1. Select whichever indexed policy has the highest vector similarity, regardless of version.
  2. Filter and rank by the currently approved document version and effective date at retrieval time.
  3. Restrict retrieval to files with the latest ingestion timestamp, even if they are still drafts.
  4. Tell the language model to prefer newer policies but supply every historical revision.

Explanation

Correct Answer

B:

Document revision and effective-date metadata make it possible to exclude superseded guidance instead of trusting similarity alone.

Incorrect Answers

A:

Similarity scores do not establish that the matched benefits policy is approved or currently effective.

C:

An ingestion timestamp is not an approval or effective-date control and may favor a newly uploaded draft.

D:

A prompt preference cannot reliably prevent an obsolete policy from entering the answer context.

Question 7

A legal research assistant must compare nuanced exceptions across long, conflicting contracts and cost is secondary to reasoning quality. What is the starting model-selection strategy?

  1. Benchmark a sufficiently capable large language model on complex contract test cases
  2. Select the cheapest classifier without testing its ability to reconcile long, conflicting contractual exceptions.
  3. Select a language model solely from a generic medical-abbreviation benchmark that contains no contract exception cases.
  4. Test a smaller model against the fixed-label accuracy target before selecting a large model

Explanation

Correct Answer

A:

Complex exception reasoning should be evaluated with a capable model against representative legal examples before selecting a cheaper alternative. This directly supports the need to handle difficult multi-document reasoning with measured quality, unlike a mechanism primarily intended to run short classifications with limited hardware and no network.

Incorrect Answers

B:

Simple classification cost is not the dominant constraint here. Benchmark a sufficiently capable reasoning model on representative complex legal cases before selecting a smaller alternative.

C:

A generic medical-abbreviation evaluation does not establish reasoning quality on contractual exceptions or conflicting clauses. The team needs representative legal cases and measured performance from a model capable of difficult multi-document reasoning.

D:

Structured narrow tasks may not require a general large model; accuracy, latency and measured spend should determine the smallest adequate choice. Its primary purpose is to pick sufficient capability for a bounded output taxonomy; the stated constraint instead demands a way to handle difficult multi-document reasoning with measured quality.

Question 8

A product team has a reference photograph and wants a new image with similar composition but different colors. What generation input should be considered?

  1. Describe the photograph only in plain text and omit the actual reference pixels from the new generation request.
  2. Use a supported reference-image-conditioned generation workflow
  3. Reuse a cached generic picture while omitting the supplied reference photograph from the new generation request.
  4. Record the photo's hash for provenance but omit the reference image from the generation inputs.

Explanation

Correct Answer

B:

Reference media can condition composition or appearance in ways that text-only instructions may not capture reliably. This directly supports the need to condition image generation on reference visual material, unlike a mechanism primarily intended to control supported video presentation parameters.

Incorrect Answers

A:

A prose description can miss the original photograph's framing and object placement. The supported image-conditioned workflow must receive the reference image to preserve composition while changing its colors.

C:

A cached asset cannot be expected to preserve the reference photograph’s requested composition while changing colors. The new generation must accept supported reference-image conditioning to maintain its visual relationship.

D:

Logging a reference photograph does not condition the new composition. Use the model's supported image/reference input to steer the generated result while separately recording provenance.

Question 9

A support team wants to group thousands of complaints by their underlying subject without predefined category labels. What analysis helps?

  1. Apply predefined priority tags as though they were confirmed complaint topics
  2. List every organization name mentioned in complaints and assume the names alone reveal unsupervised complaint themes.
  3. Supply local sentence context and domain-specific disambiguation criteria
  4. Apply topic discovery or semantic clustering to complaint texts

Explanation

Correct Answer

D:

Unsupervised topical grouping reveals common themes from a large corpus. In this scenario the team must discover recurring themes without fixed labels. A feature used to capture relationships between named entities does not substitute for it.

Incorrect Answers

A:

Priority is an operational urgency label, not a discovered complaint theme. Mapping all tickets to existing priority labels would miss latent subject clusters in this unlabeled corpus.

B:

Entity extraction produces named objects, not clusters of common issues. Theme discovery requires grouping reports by semantic subject without predefined labels.

C:

Ambiguous terms require surrounding evidence and taxonomy rules to resolve the intended sense. That is relevant to disambiguate entity meanings using context, but the present scenario requires an ability to discover recurring themes without fixed labels.

Question 10

A compliance answer must identify which revision of an indexed handbook supplied each retrieved passage, even after later edits. Which ingestion metadata provides defensible citation lineage?

  1. Store a stable source-document ID, revision or version timestamp, and original content hash with each indexed passage.
  2. Use a new random index ID each time a handbook revision is ingested and discard the original source reference.
  3. Store a chunk offset but discard the handbook's version, effective date and source content digest.
  4. Extract and normalize document text with source identifiers before indexing

Explanation

Correct Answer

A:

A reliable grounding index must retain which authoritative file and revision produced a passage. A source ID, revision identifier and content digest provide reproducible evidence when later editions differ. This is a provenance problem, separate from idempotent index upserts and stale-document deletion. Version metadata and a source content digest allow downstream citations to point to the exact handbook revision rather than only an ambiguous title.

Incorrect Answers

B:

Random IDs without a preserved source reference break source-to-index lineage and make revision audit difficult.

C:

A chunk offset locates text within a document but does not identify which revision contained it. Provenance requires stable source and revision identifiers.

D:

Normalized text and provenance enable later retrieval and citation of relevant policy passages. This does not provide the required ability to preserve source revision and document-level provenance for indexed passages.

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