AI-103 Plan and manage an Azure AI practice questions
7-day money-back guarantee — full refund within 7 days of purchase if you've completed under 20% of the questions. See pricing →
Certifications Tools Flashcards Career Paths Exam Guides Blog Pricing For Teams About

Language

✓ EnglishDeutschEspañolFrançaisPortuguês
Check readiness — free →

AI-103 Plan and manage an Azure AI solution: 364 practice questions

AI-103 364 questions 12 shown free

12 of the 364 Plan and manage an Azure AI solution questions in the Certsqill AI-103 bank, shown in full below. Each one carries an explanation for every option, not just the correct one — the wrong answers are where the marks go.

Preparing for AI-103? Take the free 5-min readiness check →

1. Create a hosted agent that runs custom orchestration code: Which implementation should you choose?

Medium
Your company needs a Microsoft Foundry agent that executes custom orchestration code while Foundry manages its runtime. A current project, compatible model, and supported region are available. Which implementation should you choose?
  1. Use Content Understanding to extract requirements into structured fields, without hosting agent code.
    Content Understanding analyzes supported content into structured results; it does not itself provide hosted agent orchestration.
  2. Create a prompt agent with instructions and configured tools, without a custom orchestration-code runtime.
    Prompt agents use instructions and tools, but this choice does not provide the required custom orchestration-code runtime.
  3. Create a hosted agent that runs custom orchestration code while Foundry manages its runtime hosting. ✓
    Hosted agents are intended for custom code and orchestration with managed hosting, matching both stated requirements.
  4. Use RAG to retrieve orchestration logic from business documents at run time.
    RAG supplies external information as model context; it is not a custom-code orchestration runtime or hosting approach.
The trap
Choose hosted agents when custom code and managed hosting are both required; choose prompt agents for configuration-driven behavior.

Hosted agents fit the explicit need for custom code, orchestration, and managed hosting.

2. Select a supported model and configure it as the prompt: What should the developer do first?

Hard
A support team creates a current Microsoft Foundry prompt agent to answer questions from approved policy documents. Instructions, document retrieval, tool configuration, permissions, and project setup are complete, but agent creation reports that no model is configured. The required model supports the planned text interaction in the selected region. What should the developer do first?
  1. Add Content Understanding to convert policy documents into structured business records instead of configuring the model.
    Content Understanding analyzes source content into structured output, but it does not provide the model required by a prompt agent. The scenario already has document retrieval configured.
  2. Replace it with a hosted agent and add custom orchestration code for document answers.
    Hosted agents are appropriate when custom code or orchestration is required, but the scenario requires neither. This changes the runtime instead of supplying the missing model configuration.
  3. Select a supported model and configure it as the prompt agent’s response model. ✓
    A prompt agent requires a model in addition to instructions and tool configuration. Selecting the supported model resolves the stated missing prerequisite; document retrieval and permissions are already complete.
  4. Add retrieval sources so the agent can answer from documents despite lacking a configured model.
    Retrieval supplies external information for model context; it cannot replace the model that generates the agent response. Adding sources does not address the reported prerequisite.
The trap
When diagnosing agent creation, verify the model, instructions, tools, compatibility, and access before redesigning the runtime.

Configure a supported model first; prompt agents require a model, instructions, and tool configuration.

3. Require answers from retrieved passages and acknowledgment: Which targeted correction should you make?

Medium
You have an Azure subscription containing a current Microsoft Foundry project and a RAG application. Evaluation shows 92% relevant retrieval, current documents, correct access filtering, but only 61% of answers supported by retrieved passages. Which targeted correction should you make?
  1. Select a larger model while leaving retrieval and grounding instructions unchanged as the primary correction.
    Model selection should consider representative task quality, but model size alone is not the targeted diagnosis. The evidence identifies insufficient answer support, requiring grounding behavior and abstention rules.
  2. Require answers from retrieved passages and acknowledgment of insufficient evidence. ✓
    Retrieval, freshness, and access are performing well; the failure is answer generation. Prompt-agent instructions can require grounded answers and abstention when evidence is insufficient.
  3. Replace the retrieval index with Content Understanding fields for every document as the primary fix.
    Content Understanding can extract structured content, but the evaluation already shows relevant retrieval. Replacing the index does not directly correct unsupported answer generation and adds an unnecessary processing change.
  4. Rebuild document permissions and ingestion before addressing unsupported answer generation in the app.
    Permissions and freshness are already reported as correct. Rebuilding them addresses neither the measured generation failure nor the requirement to diagnose retrieval and answer generation separately.
The trap
When retrieval metrics are healthy but answers lack support, adjust grounding and abstention behavior before changing indexes or source processing.

Retrieval is relevant and current, so correct the answer-generation behavior with explicit grounding and abstention instructions.

4. Configure a Content Understanding schema and map validated: What should you implement?

Medium
Your company successfully prototypes extracting invoice fields from documents. For a reproducible production rollout, the application must create business records from structured extraction results. Current Foundry resources and required service access are available. What should you implement?
  1. Create a prompt agent with invoice instructions and keep extracted output only in conversation history, without creating business records.
    This changes content representation or processing, but omits a step explicitly required by the scenario. A prompt agent is selected for model, instructions, and tools; an analyzer does not itself update the target business database.
  2. Add RAG and generate invoice records from retrieved document context.
    RAG supplies contextual information but does not replace schema-based extraction or guarantee that generated records are correct.
  3. Configure a Content Understanding schema and map validated fields into the business-record workflow. ✓
    Content Understanding produces structured fields according to an analyzer schema; application logic must use those fields to create records.
  4. Configure Markdown output and send retrieved passages directly into the business-record workflow.
    Markdown is useful for downstream retrieval, but passages are not the structured application fields required for business records.
The trap
Match the output format to the workflow: structured fields for records, Markdown for retrieval, and explicit integration for database updates.

Use a Content Understanding analyzer for schema-aligned extraction, then explicitly integrate its structured output with record creation.

5. Configure retrieval over accessible: What prerequisite should the developer add?

Medium
A current Microsoft Foundry prompt agent must answer support questions using frequently updated policy documents and acknowledge when evidence is insufficient. The project, model, instructions, tool configuration, permissions, and supported text modality are available. The agent gives generic answers because the policy documents are not available as model context. What prerequisite should the developer add?
  1. Move prompt logic to a hosted agent leaving policy retrieval unconfigured.
    Hosted agents provide managed hosting for custom code and orchestration; hosting does not automatically retrieve the organization’s documents or ground responses.
  2. Create a Content Understanding analyzer for structured policy fields.
    Content Understanding can analyze documents into structured content, but an analyzer alone does not make that content available for agent retrieval or update the business workflow.
  3. Add detailed instructions telling the model to recall current policies and cite evidence it cannot access.
    Instructions can define behavior and an insufficient-evidence response, but cannot provide unavailable, changing policy content. This fails the external-knowledge prerequisite.
  4. Configure retrieval over accessible, indexed policy documents so relevant evidence enters agent context. ✓
    Retrieval supplies relevant external documents as model context, enabling grounded answers without retraining. Document freshness, relevance, access, and answer support still require inspection.
The trap
Separate agent behavior configuration from knowledge availability; diagnose retrieval and answer-generation failures independently.

The missing prerequisite is a retrieval path over accessible policy documents, allowing relevant evidence to be supplied to the prompt agent.

6. Add instructions requiring acknowledgment of insufficient: Which targeted correction addresses that behavior?

Hard
You have an Azure subscription with a current Microsoft Foundry project containing a prompt agent and retrieval tool. Model support and document access are verified. Testing shows current policy answers cite relevant documents, superseded-policy questions retrieve older documents, and unsupported questions receive confident answers with invented citations. The agent must acknowledge insufficient evidence instead of making unsupported claims. Which targeted correction addresses that behavior?
  1. Configure a Content Understanding analyzer to extract policy fields leaving retrieval and response instructions unchanged.
    Content Understanding can structure document content, but an analyzer does not update the target business database or address unsupported answer behavior.
  2. Replace the model with a larger model because increased size should prevent unsupported policy responses.
    Model selection should consider task quality, modalities, latency, and cost; a larger model does not supply the required evidence-handling rule.
  3. Add instructions requiring acknowledgment of insufficient evidence when retrieved content cannot support an answer. ✓
    An explicit insufficient-evidence stopping condition directly addresses fabricated answers; freshness and retrieval quality can be evaluated separately.
  4. Retrain the model on policy documents so retrieval is no longer needed for policy answers.
    RAG provides current external information without retraining, and retraining does not establish the required behavior when evidence is absent.
The trap
Separate retrieval failures from answer-generation failures; require an explicit insufficient-evidence behavior.

Add an explicit prompt-agent instruction to acknowledge insufficient evidence rather than fabricate an answer or citation.

7. Create a representative test set: Which step best satisfies these requirements?

Medium
Your company has a successful current Microsoft Foundry RAG prototype for internal policy questions. Before production rollout, the team must reproduce quality checks, detect stale or inaccessible documents, and distinguish retrieval failures from answer-generation failures. The model, region, and data permissions are already configured. Which step best satisfies these requirements?
  1. Create a representative test set, then measure retrieval relevance and answer support separately during each rollout. ✓
    This creates a repeatable quality baseline, diagnoses retrieval separately from generation, and checks whether answers are supported by retrieved evidence.
  2. Convert policy documents into structured fields and validate their schema instead of testing retrieval and answer support.
    Structured fields can support application records, but this omits the required evaluation of retrieval, freshness, access, and answer support.
  3. Move orchestration into a hosted agent and validate deployment health instead of retrieval and answer quality.
    Hosted agents provide custom code and orchestration, not guaranteed truth, freshness, authorization, or a substitute for separate evaluation.
  4. Reindex all documents and rely on generated citations as proof that every answer is factually supported.
    Reindexing can refresh retrieval, but citations do not prove truth or answer support; freshness, relevance, access, and evidence require inspection.
The trap
For RAG rollout questions, distinguish retrieval quality from answer support and verify evidence rather than trusting citations.

Use a repeatable evaluation set and independently assess retrieval quality and answer support before each rollout.

8. Add app integration that maps the structured output: What should the developer do next?

Easy
A claims application uses Content Understanding to extract policy number, claimant, and loss date from uploaded documents. Analyzer events show successful structured output, but the target business database remains unchanged. What should the developer do next?
  1. Add retrieval indexing so the extracted fields become available to the target business database.
    Retrieval can support downstream search and grounding, but indexing does not itself write structured analyzer results into the target database.
  2. Change the analyzer output to Markdown so the target business database can consume the extracted records.
    Markdown can support downstream retrieval, whereas structured fields better support application records; changing format does not create persistence.
  3. Add app integration that maps the structured output and writes the fields to the target business database. ✓
    The analyzer completed successfully; the missing step is downstream handling that maps its fields and persists them in the database.
  4. Update prompt-agent instructions to save analyzer fields without implementing downstream database integration.
    Instructions and tool configuration guide prompt agents, but they do not update a target database without an implemented integration path.
The trap
When analyzer events succeed but records are missing, inspect the handoff and persistence path before changing the analyzer or retrieval design.

Successful analyzer output proves extraction, not persistence. Implement the application integration that maps and writes the structured fields.

9. Configure the prompt agent to state when evidence is: Which targeted correction addresses the reported defect?

Medium
You have an Azure subscription containing a current Microsoft Foundry project, a deployed text-capable model, a configured retrieval index, and a Content Understanding analyzer. Evaluation shows that retrieval usually finds relevant policy documents, answers sometimes invent citations when evidence is missing, and invoice fields are extracted correctly. Which targeted correction addresses the reported defect?
  1. Change the analyzer output from structured invoice fields to Markdown so the agent can verify citations.
    Markdown can support downstream retrieval, but changing invoice extraction format does not correct fabricated citations when retrieval evidence is insufficient.
  2. Replace the deployed model with a larger model because model size alone should determine citation reliability.
    Model selection should consider task quality, modalities, latency, and cost—not size alone. This does not specifically correct unsupported citation behavior.
  3. Configure the prompt agent to state when evidence is insufficient instead of fabricating citations, using explicit stopping instructions. ✓
    This directly addresses unsupported answers while preserving successful retrieval and extraction. Stopping behavior should be explicit.
  4. Move the existing prompt and retrieval logic into a hosted agent without changing its evidence checks.
    Hosted agents provide managed hosting for custom code and orchestration; hosting does not automatically validate evidence or citations.
The trap
When retrieval succeeds but unsupported answers remain, inspect evidence handling and stopping instructions before changing models or services.

Use explicit prompt-agent instructions to acknowledge insufficient evidence and stop rather than fabricate citations.

10. Record and reproduce the model: Which action should the developer take first?

Hard
Your company has a successful prototype prompt agent in a current Microsoft Foundry project. The agent must be rolled out reproducibly to another environment with the same supported model, instructions, tools, explicit stopping conditions, and representative-task acceptance tests. No custom code or document ingestion is required. Which action should the developer take first?
  1. Configure Content Understanding to write analyzed prototype conversations into the target business database.
    Content Understanding analyzes supported content into structured output, but it does not itself update a business database and is unrelated to this rollout.
  2. Add RAG and use retrieved documents as the acceptance test for every response.
    RAG grounds responses in external information but does not provide reproducible prompt-agent configuration or replace broader task evaluation.
  3. Reimplement the prototype as hosted-agent code before establishing versioned configuration and representative acceptance tests.
    Hosted agents suit scenarios requiring custom code or orchestration, which this scenario excludes; hosting alone does not ensure identical behavior.
  4. Record and reproduce the model, instructions, tools, stopping conditions, and representative-task acceptance tests. ✓
    This preserves the prompt agent’s essential configuration and verifies that rollout meets the same task-quality requirements.
The trap
For prompt-agent scenarios, preserve configuration and observable evaluation criteria before changing agent types or adding unrelated services.

Reproduce the prompt agent’s model, instructions, tools, stopping conditions, and representative-task tests, after confirming target support.

11. Verify document access: What should the developer do first?

Medium
A current Microsoft Foundry RAG agent must answer from newly uploaded company policies and acknowledge insufficient evidence. Traces show generation completed, but retrieval returned zero passages. Documents are intended to be accessible to the agent. What should the developer do first?
  1. Replace the selected model with a larger model before investigating the retrieval dependency.
    Model selection should consider representative task quality, modalities, latency, and cost, but a larger model does not resolve zero retrieved passages.
  2. Verify document access, freshness, and retrieval relevance, then repair the retrieval dependency. ✓
    Zero retrieved passages indicates a retrieval-side failure. Verify that documents are current and accessible and that retrieval finds relevant content before diagnosing generation.
  3. Configure Content Understanding to extract policy fields before checking the retrieval tool and index.
    This changes content representation or processing, but omits a step explicitly required by the scenario. Content Understanding can produce structured content, but its analyzer does not by itself update a target business database or repair this retrieval failure.
  4. Strengthen instructions requiring citations for every policy answer before fixing the missing evidence returned by retrieval or its source documents.
    Instructions cannot supply missing evidence and may encourage unsupported answers. The agent should acknowledge insufficient evidence rather than fabricate citations.
The trap
Use observable retrieval and generation events to isolate the failing stage before changing prompts or models.

Zero passages points first to retrieval or document dependencies, not answer generation.

12. Add application code to validate fields and write approved: What should you add to address the remaining risk?

Easy
You have an Azure subscription that contains a current Microsoft Foundry project and a Content Understanding analyzer. After improving its invoice analyzer, extracted fields match the required schema. The business still needs those fields available in its accounts-payable database. The analyzer is configured and the application has permission to write there. What should you add to address the remaining risk?
  1. Archive analyzer output as Markdown and omit app logic that writes validated fields to the accounts-payable database.
    This changes content representation or processing, but omits a step explicitly required by the scenario. Markdown can support downstream retrieval, but changing the output representation does not synchronize records to a database.
  2. Add application code to validate fields and write approved database records. ✓
    Content Understanding produces structured content but does not itself update the target business database. An application integration must perform that write.
  3. Expand the analyzer schema and rely on analysis completion to commit extracted records to the database.
    Schemas determine structured analyzer output; analysis completion does not perform the separate database-write integration.
  4. Add a prompt agent to retrieve analyzer output and assume each extracted field is already synchronized to the database.
    A prompt agent can use configured tools, but merely retrieving output does not update the target database or provide synchronization.
The trap
Separate content analysis from downstream persistence: an analyzer’s successful result is not a database update.

Content Understanding extracts structured content but does not update business databases; application logic must validate and persist the fields.

352 more Plan and manage an Azure AI solution questions

The remaining 352 questions in this domain are part of the full AI-103 bank — 1300 questions, every option explained. Start with the free five-minute check and see your score per domain.

Test your AI-103 readiness — free

Other AI-103 domains

Part of the Certsqill AI-103 question bank · Plan and manage an Azure AI solution · Every answer, right and wrong, comes with its own explanation.