Microsoft AI-103 Practice Test 2
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?
- Store a stable source-document ID, revision or version timestamp, and original content hash with each indexed passage.
- Use a new random index ID each time a handbook revision is ingested and discard the original source reference.
- Store a chunk offset but discard the handbook's version, effective date and source content digest.
- 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.
Question 11
An engineer sends a text-only request to a model that requires an image part for visual tasks. What must be changed?
- Change the model endpoint URL but continue sending a text-only prompt that contains no visual evidence.
- Change only the deployment identifier while continuing to send a text-only payload for the visual request.
- Submit a supported multimodal message containing image input and accompanying text
- Capture usage counts with the deployment name for per-model cost attribution
Explanation
Correct Answer
C:The message payload must include the actual visual content in a format accepted by the multimodal API. This directly supports the need to supply image evidence in model input, unlike a mechanism primarily intended to select the correct model serving API.
Incorrect Answers
A:Correct endpoint routing alone cannot supply the missing image data. The request payload must include a supported image part alongside relevant text.
B:A different deployment ID does not supply the missing image content. The accepted multimodal request must contain an image part plus accompanying text.
D:Usage telemetry tied to deployment identifiers supports spend and optimization decisions. Its primary purpose is to account for input and output model tokens; the stated constraint instead demands a way to supply image evidence in model input.
Question 12
A retailer wants customers to submit photographs and questions about damaged packaging in a single interaction. Which model capability is essential?
- Select a multimodal model that accepts images together with text
- Send only a written description to a text-and-retrieval service, never including the submitted damaged-package photograph.
- Evaluate only written medical abbreviations and use the resulting text-only model to inspect customer photographs.
- Use a text-only classifier and extract package condition solely from the customer's written question, ignoring the uploaded photograph.
Explanation
Correct Answer
A:A text-only model cannot directly inspect pixels; a multimodal input model can jointly use the submitted image and the accompanying question. The decisive difference is whether the design can interpret images and text in one request; the competing function serves to separate language generation from live authoritative data.
Incorrect Answers
B:Retrieval of product facts does not inspect the customer's actual package image. A suitable multimodal model must receive the photograph together with the user's question in the same supported request.
C:A text-only benchmark says nothing about whether a model accepts visual data. The application requires a multimodal endpoint that can process both the damaged-package photograph and accompanying customer question.
D:Text-only inference cannot inspect the damaged package shown in the customer's photo. The application needs a model endpoint that accepts visual evidence and text together.
Question 13
A brand campaign requires several seconds of synthetic moving visuals rather than a still image. What model capability is required?
- Run image-to-image editing on a single frame and deliver the edited still as though it contained continuous motion.
- Reuse one approved still-image asset throughout the entire clip and claim it supplies temporal motion without video generation.
- Select an available text-to-video generation capability and suitable output constraints
- Repeatedly submit a refused image prompt instead of choosing a model that generates motion.
Explanation
Correct Answer
C:Video generation produces a temporal sequence; a still-image API will not produce moving output on its own. The decisive difference is whether the design can create moving video clips from prompts; the competing function serves to condition image generation on reference visual material.
Incorrect Answers
A:Editing one still image does not generate a temporal video sequence. The campaign requires a video-capable generation operation with the intended motion and duration.
B:Caching a still frame does not create a sequence with genuine moving visual content. The deliverable calls for a supported text-to-video workflow with appropriate duration and output constraints.
D:Retrying an unrelated safety response cannot create temporal video frames. A supported video-generation workflow is required for moving output.
Question 14
A support model correctly extracts names but confuses the payer and recipient in a long invoice paragraph. What output design helps?
- Extract typed entity relationships with explicit payer and recipient roles.
- Return the two recognized names as an unordered array of persons.
- Apply sentiment analysis to the entire invoice paragraph before extracting names.
- Tag both people with the same generic PERSON type and omit relation edges.
Explanation
Correct Answer
A:Recognizing two entity strings is insufficient; role labels express who performed or received each transaction.
Incorrect Answers
B:An unordered entity list retains the strings but loses which party pays and which party receives.
C:Sentiment classification does not assign transaction roles to the recognized entities.
D:Identical entity labels cannot disambiguate payer from recipient in a long transactional sentence.
Question 15
A retrieval pipeline loads one scanned PDF as a single bitmap and search yields no text for an important clause. What preprocessing is needed?
- Index the pixel dimensions and filename of the scan without extracting characters.
- Run vector similarity over the document filename alone.
- Run appropriate OCR and page layout extraction before text indexing.
- Enable semantic reranking on the empty text index.
Explanation
Correct Answer
C:A bitmap-only scan cannot satisfy lexical search until text is recovered from the image and its page structure.
Incorrect Answers
A:Image metadata cannot provide the clause text required for lexical retrieval.
B:Embedding the filename cannot recover words visible only inside scanned page images.
D:A reranker cannot rank missing extracted clause text and cannot replace OCR and layout parsing.
Question 16
An app's completion request fails because the client points at a project endpoint rather than the required model invocation route. What should be checked?
- Keep calling the project's configuration API with a different retry delay instead of using a supported model inference route.
- Use the correct service endpoint and API contract for the chosen model deployment
- Add a response JSON schema to the same project-management request instead of moving the call to the model inference route.
- Retry the same project-management endpoint with a deployment ID but no supported inference route.
Explanation
Correct Answer
B:Foundry project management routes and model inference endpoints are not necessarily interchangeable. The decisive difference is whether the design can select the correct model serving API; the competing function serves to manage model request concurrency and throttling.
Incorrect Answers
A:Retry timing does not change which API operation receives the request. The application must direct completions to the model-serving endpoint and contract for its actual deployment.
C:A schema affects validation or generation format after a supported model request has reached the serving API. It cannot repair a request sent to an endpoint that does not handle inference.
D:A deployment name does not fix an incorrect endpoint contract. Use the model-serving SDK or inference API for the selected deployment.
Question 17
A deployment in region A succeeds but the identical Foundry model SKU is absent in region B. Before switching the release region, what should the team establish?
- Assume a model available in one Azure region is automatically available everywhere.
- Check the selected model/version's regional availability and deployment type in the Foundry catalog.
- Change only the client's endpoint hostname and reuse the region A capacity allocation.
- Check whether the subscription has sufficient storage-account quota in region B.
Explanation
Correct Answer
B:Model availability and supported deployment types can vary by Azure region. The subscription quota and deployment configuration must be checked in the target region before cutover.
Incorrect Answers
A:Regional catalogs and supported deployment types differ; successful deployment in region A proves nothing for B.
C:A new hostname does not allocate the chosen model and deployment type or quota in region B.
D:Storage quota does not establish model-version availability or Foundry inference throughput capacity.
Question 18
A customer requests a product photo where only the background changes while the device label stays intact. What editing input should be provided?
- Regenerate the entire photograph from a general description of the device.
- Supply a mask identifying the editable background and preserve the protected product region.
- Use a mask that covers the label and device but excludes the background.
- Crop out the original background and discard the unedited product without compositing.
Explanation
Correct Answer
B:Mask-guided editing directs generation to the intended pixels; unrestricted regeneration can alter factual details on the product.
Incorrect Answers
A:Full regeneration can change label text or device features that the customer explicitly requires preserving.
C:This reverses the intended edit region and risks replacing the protected label rather than the background.
D:A crop alone does not produce a faithful edited background while retaining the original device label.
Question 19
A news summarizer must distinguish the person mentioned from the person who performed an action. What representation helps?
- Extract entities together with their contextual roles or relations
- Return person names alone without the actions or grammatical roles connecting them
- Classify the article's overall topic without representing which named person performed the described action.
- Extract named entities and domain-defined identifier fields
Explanation
Correct Answer
A:Entity names alone do not describe who did what to whom in a sentence. This resolves the specific need to capture relationships between named entities, while a measure that helps disambiguate entity meanings using context addresses a different requirement.
Incorrect Answers
B:An entity list identifies people but cannot distinguish the actor of an event from its subject. The requirement is relational attribution in a particular news sentence.
C:Topic classification does not express agent, actor or recipient relationships. The task requires typed roles or relations between extracted entities.
D:Entity extraction identifies spans that represent people, organizations and specific domain values. The function serves to identify entity spans and policy identifiers, whereas the actual requirement is to capture relationships between named entities.
Question 20
Large manuals are stored as single index documents and retrieval supplies irrelevant chapters. How should they be partitioned?
- Index each complete manual as one vector, then request fewer results without splitting chapters
- Split documents into semantically coherent chunks with stable source offsets
- Keep each 200-page manual in one giant index entry but assign a stable document ID to it.
- Track source deletions and remove or mark deleted index entries
Explanation
Correct Answer
B:Well-designed chunks improve retrieval specificity while retaining citation lineage. The decisive difference is whether the design can index targeted passages rather than whole manuals; the competing function serves to prepare document text and identity for retrieval.
Incorrect Answers
A:Changing the top-k count does not create chapter-level retrieval units; irrelevant passages remain embedded in an oversized document. Chunking should respect section boundaries and preserve provenance.
C:Stable identifiers help updates, but whole-manual chunks still hide the relevant passage among unrelated chapters. Segment by coherent sections with offsets and provenance.
D:Ingestion must reconcile removed source content so responses do not cite obsolete files. That is relevant to synchronize indexed deletions, but the present scenario requires an ability to index targeted passages rather than whole manuals.