Azure AI Apps and Agents Developer Associate (AI-103) - Full Practice Exam 1

1. A business requires evidence of which agent tool changed a record and which user authorized it. What should auditing capture?
2. A claims workflow receives multiple document types. Each type needs a different set of extracted fields. Which design fits?
3. A client submits an image-generation request with an unsupported size parameter. What should the integration do?
4. A contact-center application needs to locate and redact supported sensitive information in transcripts before broader analysis. Which capability should be evaluated?
5. A document workflow must identify organizations, dates, and locations before applying business rules. Which text analysis capability fits?
6. A Foundry-related supported service must be accessed only through the organization's private network. Which design elements must be evaluated together?
7. A long meeting transcript needs a concise action-oriented summary. The engineer wants to prevent unrelated or unsupported commitments from appearing. Which evaluation is most important?
8. A marketing application must create a new image from a textual brief. Which deployment capability should be selected?
9. A model deployment is updated. The existing agent still returns text, but tool arguments now frequently violate the expected contract. What should happen?
10. A model must answer from frequently updated private operating procedures without retraining whenever a document changes. Which architecture is most appropriate?
11. A multimodal prompt asks the model to compare two product photographs. Only one image is included in the request. What should the client correct?
12. A new prompt improves average answer relevance but increases unsafe tool calls. What should the release decision consider?
13. A prompt template embeds retrieved text between delimiters and explicitly identifies it as reference material. Why is this useful?
14. A public assistant needs detection of harmful content categories in user input and generated responses. Which capability should be evaluated as part of the safety design?
15. A RAG assistant cites relevant documents but includes a claim that none of them supports. Which evaluation dimension most directly targets this problem?
16. A release candidate answers normal questions well but leaks protected facts when given adversarial instructions. What should happen before production rollout?
17. A review-processing API must distinguish overall opinion and opinions about specific aspects where supported. Which capability is most relevant?
18. A search index uses embeddings from one model, but query vectors come from an incompatible embedding space. Which corrective action is appropriate?
19. A support application must translate written messages into another language while retaining their meaning. Which service capability directly fits?
20. A support corpus contains exact product codes and conceptual descriptions. The team wants both lexical precision and semantic similarity. Which retrieval approach should it evaluate?
21. A system must extract structured information from supported audio and video, in addition to documents. Which service should the engineer evaluate for a unified schema-oriented analysis workflow?
22. A team changes models but keeps the same number of requests. Costs increase sharply because prompts and answers are much longer. What should its monitoring include?
23. A team wants to reproduce which model, prompt version, retrieval configuration, and dataset produced an evaluated result. What should its delivery process retain?
24. A tool response includes text instructing the agent to reveal credentials. How should the orchestration layer treat it?
25. A visual question-answering application sends an image and asks for the serial number visible on a damaged label. The characters are unclear. What response policy should it follow?
26. A voice assistant needs to accept spoken questions and return spoken answers. Which two transformations are required around a text-based reasoning component? Select TWO.
27. A workflow delegates a retrieval subtask to one specialist agent and a calculation subtask to another. What should the coordinator preserve?
28. An agent has a shipment lookup tool and a refund tool. A user asks only where a package is. Which tool-access design reduces unintended business changes?
29. An agent may modify customer accounts. Security wants to limit consequences of an incorrect model decision. Which control is most appropriate?
30. An agent must remember a user's preference across turns but must not mix records between users. Which design is appropriate?
31. An agent repeatedly alternates between two tools without completing the task. Which safeguard should be added?
32. An agent tool charges a customer. A network timeout occurs after the backend may already have committed the charge. What should the integration support before retrying?
33. An AI support application retrieves obsolete policy passages even though the generator is functioning. Which component should be investigated first?
34. An AI workflow is moved from test to production. Which deployment practice best supports safe release?
35. An AI workflow needs a generated video, while its current deployment accepts images and returns text. What should the engineer verify?
36. An application must accept customer images and return grounded language answers about their contents. Which TWO model-selection checks are necessary? Select TWO.
37. An application needs a stable JSON response contract for downstream validation. Which implementation is strongest?
38. An application produces captions for product images. Which evaluation best determines whether the captions are suitable?
39. An application receives repeated rate-limit responses after adding concurrent clients. What should the engineer investigate first?
40. An Azure application calls an eligible AI service using a key embedded in its source. The organization wants keyless authentication where supported. Which approach fits?
41. An engineer must choose a model for extracting structured facts from short messages. A small model meets quality targets with lower latency and cost. What is the best selection approach?
42. An evaluation dataset includes only ideal, clearly phrased queries. What addition most improves confidence before public deployment?
43. An extraction result will automatically update a financial system. A required amount is uncertain or missing. What should the application do?
44. An extraction schema expects a date and numeric total, but the returned values are malformed. Which boundary should reject or route the result before persistence?
45. An image-editing workflow should alter one specified region while retaining the surrounding image. Which supported model feature should be evaluated?
46. Content analysis is asynchronous. A worker receives an operation reference and then crashes. Which workflow design best supports safe recovery?
47. Long policy files are indexed as one vector each, and specific answers are difficult to retrieve. Which improvement should be tested?
48. Reviewers need to verify which source passage supports a structured field. Which analyzer output should be preserved where supported?
49. The assistant's response is consistently too verbose. Which first adjustment best targets the behavior without changing the business task?
50. Users may access different policy documents. How should a RAG application prevent retrieval of documents the caller is not authorized to see?

Comments — 0