AWS Generative AI Developer Professional (AIP-C01) - Full Practice Exam 1

1. A Bedrock agent must propose refunds but the application must authorize and execute them through an existing transaction service. The team wants to inspect the selected action before fulfillment. Which integration is most appropriate?
2. A Bedrock invocation fails with AccessDenied before generation begins. What is the most appropriate investigation?
3. A business wants to customize a model using confidential documents. What should be established before submitting the training job?
4. A chat application needs to retain useful conversation context but has a fixed token budget. What is the best state-management design?
5. A chat application reuses one Bedrock agent session ID for all logged-in users. What is the most important correction?
6. A company adds Bedrock Guardrails to a customer assistant. Which statement correctly describes the remaining security responsibilities?
7. A company discovers that a third-party document used in its AI corpus has unclear usage rights. What is the appropriate governance response?
8. A company evaluates a new foundation model for a customer-facing assistant. Which two criteria should drive selection beyond a public benchmark score? Select 2 answers.
9. A company must track the source, license, approval, and preprocessing history of documents used in an AI system. What should the data pipeline preserve?
10. A company needs a model to consistently produce its approved report style. Facts come from a fresh retrieval corpus, but few-shot prompting still misses the required style. Which next experiment is most appropriate?
11. A company requires inference processing only in approved geographic locations. It wants to use a cross-Region inference profile for capacity. What must be checked before enabling it?
12. A company wants to reuse and version prompt templates across deployments. Changes must be reproducible and evaluated before release. Which capability should be considered?
13. A corpus sync recomputes embeddings for every unchanged document each day. What optimization best preserves correctness while reducing unnecessary work?
14. A customer assistant should begin showing text quickly while a model generates a long answer. Which design best improves perceived responsiveness without claiming faster total generation?
15. A customer-facing assistant is tested only with polite ordinary questions. What additional evaluation is needed before release?
16. A document ingestion job is interrupted halfway through a large corpus. The team wants safe reruns without duplicating all indexed content. Which design is best?
17. A document is deleted from the source repository because access was revoked. The assistant still retrieves its old chunks. What is the best remediation?
18. A frontend streams model output but treats a dropped connection as a completed successful answer. What integration change is most appropriate?
19. A grounded-answer check flags a response as supported by its retrieved passage, but that passage belongs to another tenant. What is the correct assessment?
20. A helpdesk corpus contains exact product IDs as well as paraphrased troubleshooting descriptions. Semantic search misses some exact identifiers. Which retrieval change should be evaluated?
21. A knowledge base service role can read its S3 source, but an OpenSearch Serverless-backed ingestion fails with access errors. Which layers should be checked?
22. A legal policy assistant retrieves versions from several years and answers with a superseded policy despite a newer document being present. Which improvement best addresses this?
23. A long-context assistant regularly sends the same large reference prefix followed by a short changing question. What optimization should be evaluated on a supported model?
24. A model-as-judge evaluation gives a candidate model high scores, but domain experts find serious factual errors. What is the best next step?
25. A model-backed API frequently runs longer than the browser request budget. Users must retrieve completed results later. Which architecture best meets the requirement?
26. A model generates JSON for a downstream workflow. Some responses contain invalid types or extra prose. What should the integration enforce?
27. A model produces correct short answers but is prompted to explain every simple extraction in several paragraphs. Cost and completion latency are excessive. Which change should be tested first?
28. A multi-step AI workflow calls extraction, a business rules service, and summarization. One step fails transiently. What is the best orchestration design?
29. A multi-tenant assistant uses one knowledge store. Tenant metadata exists on documents, but the client supplies the filter value. What is the best isolation design?
30. A multilingual assistant retrieves English passages well but fails on supported non-English customer questions. Which two actions best investigate the gap? Select 2 answers.
31. A new embedding pipeline normalizes document text differently from query text and retrieval quality drops. What should be investigated?
32. A nightly workload evaluates thousands of independent documents and does not require immediate responses. Which inference mode should be evaluated for supported models?
33. A product team uses the Converse API to reduce custom integrations across foundation models. Which assumption must still be validated for each selected model?
34. A production Bedrock Flow must be updated without silently changing every caller to an untested draft. Which deployment approach is appropriate?
35. A prompt release improves the overall score but causes a large regression for non-English support requests. Which release decision is most appropriate?
36. A RAG application retrieves a relevant heading but misses the paragraph defining its exception. The text was split into tiny chunks without overlap or document structure. Which change should be tested first?
37. A RAG assistant answers incorrectly. The generator produces a correct answer when given the right source passage manually. What should the team investigate first?
38. A RAG corpus contains duplicate copies of the same policy in several folders. The top results repeat those passages and crowd out complementary evidence. Which ingestion improvement is best?
39. A RAG system must support citations that users can verify. Which answer-generation design is most appropriate?
40. A RAG team wants direct control over retrieval, reranking, prompt construction, and answer generation rather than one combined call. Which API pattern best fits?
41. A response cache reduces model calls but occasionally returns one user private answer to another user. Which correction is essential before retaining the optimization?
42. A retrieval call returns ten passages, but only two fit the generation budget. What should the context builder do?
43. A retrieved support document contains instructions telling the assistant to ignore its system rules and send private data to an external URL. Which response best preserves the trust boundary?
44. A sensitive-data filter masks personal information in a final answer, but the original prompt is already stored in invocation logs. What additional control is needed?
45. A stable high-volume inference workload needs predictable capacity. The selected model supports Provisioned Throughput. What should the team assess before purchasing it?
46. A support assistant must answer from product manuals updated daily and show the source passages. The model already follows the desired response format. Which approach best provides freshness and provenance with minimal retraining?
47. A team changes its embedding model to one with a different vector dimension. What must be planned for the existing vector index?
48. A team changes the generator, embedding model, and chunking policy simultaneously and sees worse answers. What practice would make future regressions easier to diagnose?
49. A team chooses a model with a large context window and plans to send its entire document archive for every question. Which concern should the architect raise?
50. A team encrypts its knowledge store but allows every application user to retrieve all documents. Which requirement remains unmet?
51. A team must choose between a general foundation model and a smaller task-adapted model for high-volume extraction. Which experiment best supports the decision?
52. A team prepares examples for model customization. The evaluation set contains near-duplicates of training examples. Why should this be corrected?
53. A team reports an improved AI assistant using a test set repeatedly inspected during prompt tuning. What is needed for a credible final quality estimate?
54. A team routes all tasks to its largest model. Many requests are simple classifications, while a small subset requires difficult reasoning. Which optimization should be evaluated?
55. A team stores long-lived agent memory derived from user conversations. What two controls are appropriate for privacy and correctness? Select 2 answers.
56. A team wants to apply the same supported Bedrock content safeguards to text from a non-Bedrock model. Which API capability should it evaluate?
57. A text-to-SQL assistant should answer analytics questions but must never modify production data. Which two controls are appropriate? Select 2 answers.
58. A tool-enabled model emits arguments for a customer lookup. The request contains a customer ID different from the authenticated user. What should the tool executor do?
59. An agent can cancel customer subscriptions. The product requires explicit customer confirmation for that action before the function is invoked. Which capability should be configured?
60. An agent executor only needs to read one secret and invoke one internal API. What is the best runtime permission design?
61. An agent invokes the wrong operation because two tools have vague overlapping descriptions and nearly identical parameter names. What should be improved first?
62. An agent needs an API token to call an approved external system. Where should token handling occur?
63. An agent retries a create-ticket tool after a timeout. The first call may already have created the ticket. Which design best prevents duplicate tickets?
64. An agent selects the correct tool but sends the wrong date parameter. Which evidence most directly helps diagnose the error?
65. An AI application can browse external URLs returned in model output. Security wants to prevent arbitrary internal-network access and data exfiltration. Which control is most important?
66. An AI application must call Bedrock from private subnets without public internet egress. Which approach supplies the private service path?
67. An AI platform compares two models. One has cheaper tokens but needs more retries and manual correction. Which metric best reflects the business efficiency trade-off?
68. An AI workflow must call several downstream services. One service has a strict concurrency limit and cannot absorb the model agent burst rate. Which design best protects the dependency?
69. An application enables full model invocation logging for a regulated workload. Which two operational controls should accompany that choice? Select 2 answers.
70. An assistant must read scanned PDF forms with tables and answer questions about their contents. Plain text extraction loses the table relationships. Which improvement is most appropriate?
71. An assistant must refuse to answer when its approved corpus has insufficient evidence. Which two evaluation practices best test this behavior? Select 2 answers.
72. An assistant proposes a high-impact business decision using incomplete evidence. The application policy requires human review. Which workflow is appropriate?
73. An organization changes its prohibited-content policy. The deployed assistant uses an older guardrail version even though a new draft exists. What should the release process do?
74. Bedrock requests receive transient throttling during bursts. Which two changes most appropriately improve stability? Select 2 answers.
75. The first token arrives quickly, but long responses finish slowly. Which measurements best distinguish startup from generation throughput?

Comments — 0