Foundation Model Integration, Data Management, and Compliance

1. A team must validate that a proposed GenAI architecture can meet its latency and quality targets before committing to full development. Which approach is appropriate?

1. A Build the full application and measure it once the implementation is complete.
2. B Build a narrow proof of concept on Amazon Bedrock that exercises the critical path and measure it against the targets.
3. C Estimate the latency from the model's published token generation rate.
4. D Review published benchmarks for the candidate models and infer the outcome.

2. An organization must ensure GenAI workloads are designed consistently against recognised architectural guidance. Which solution meets these requirements?

1. A Review each workload informally at the weekly architecture meeting.
2. B Review each workload against the team's own checklist written at project start.
3. C Review each workload with the AWS Well-Architected Tool using the Generative AI Lens and track the improvement plan.
4. D Review each workload only when an incident indicates a design problem.

3. Several teams are each building their own retrieval and prompt handling code for similar GenAI applications. Which approach is appropriate?

1. A Publish standardized components for retrieval, prompt handling, and guardrail application that each team consumes.
2. B Allow each team to build its own implementation so it fits their use case exactly.
3. C Document the recommended approach in a shared wiki for teams to follow.
4. D Have one team build the application and the others adopt it wholesale.

4. An application must be able to switch foundation model providers without changing application code. Which solution meets these requirements?

1. A Place the model identifier and its parameters in AWS AppConfig and resolve them at invocation so a change is a configuration deployment.
2. B Store the model identifier in an environment variable set at deployment time.
3. C Hardcode the model identifier and redeploy the application when it changes.
4. D Maintain a separate build of the application for each candidate model.

5. A GenAI application must continue operating when its primary model's Region experiences a service disruption. Which solution meets these requirements?

1. A Provision additional throughput in the primary Region ahead of any disruption.
2. B Configure cross-Region inference, and implement graceful degradation in the application for the period before failover completes.
3. C Cache every previous response and serve those during the disruption.
4. D Increase the retry count against the primary Region until the service recovers.

6. A domain-adapted model produced with a parameter-efficient technique must be versioned and rolled back if a deployment degrades quality. Which solution meets these requirements?

1. A Keep the adapter versions in Amazon S3 and record which is current in a document.
2. B Overwrite the deployed adapter with each new version and retrain if quality degrades.
3. C Fine-tune the base model again whenever the deployed version underperforms.
4. D Register each adapter version in the SageMaker AI Model Registry and deploy by reference so a previous version can be restored.

7. Input documents arrive as scanned PDFs, recorded audio, and structured tables, and all must be prepared for a multimodal application. Which solution meets these requirements?

1. A Send each file type directly to the model and rely on it to interpret the format.
2. B Convert every file type to plain text with a single generic parser.
3. C Store the files unprocessed and retrieve them at query time.
4. D Route each type through the appropriate extraction service before a common formatting step prepares the content for the model.

8. A data validation step must stop malformed records reaching a fine-tuning pipeline and record why each was rejected. Which solution meets these requirements?

1. A Sample the records manually before each pipeline run.
2. B Load the records and inspect the resulting model's behaviour for signs of bad data.
3. C Evaluate the records against a data quality ruleset in the pipeline and route failures to a quarantine location with their rule violation.
4. D Publish a schema document for the producing team to follow.

9. User-supplied text arrives with inconsistent casing, encoding artifacts, and truncated sentences, and response quality is poor as a result. Which solution meets these requirements?

1. A Increase the maximum output tokens so the model can elaborate on unclear input.
2. B Normalize the input before invocation, repairing encoding and casing and flagging truncated content.
3. C Instruct the model in the system prompt to ignore formatting problems.
4. D Increase the model's temperature so it compensates for the inconsistent input.

10. A knowledge base must organise documents so that retrieval can be restricted to a product line and a document type simultaneously. Which solution meets these requirements?

1. A Create a separate knowledge base for each combination of product line and document type.
2. B Include the product line and document type in the chunk text so the model can infer them.
3. C Retrieve from the whole corpus and instruct the model to consider only matching documents.
4. D Attach product line and document type as metadata to each chunk and apply both as filters at retrieval.

11. A vector store must remain current as source documents are added, updated, and deleted throughout the day. Which solution meets these requirements?

1. A Retrieve more candidates so superseded chunks are outweighed by current ones.
2. B Add new documents continuously and leave superseded chunks in the index.
3. C Rebuild the entire index each night from the current source set.
4. D Detect source changes and run incremental ingestion that upserts new and modified chunks and removes deleted ones.

12. A vector index serving several distinct subject domains returns results from the wrong domain for ambiguous queries. Which solution meets these requirements?

1. A Increase the similarity threshold so weaker cross-domain matches are excluded.
2. B Increase the number of candidates retrieved so the correct domain appears somewhere.
3. C Re-embed the corpus with a larger embedding model.
4. D Segment the index by domain, or filter on a domain attribute derived from the query context, so retrieval is scoped before ranking.

13. Retrieval performs poorly for queries that use internal product codes absent from the embedding model's training data. Which solution meets these requirements?

1. A Implement hybrid search combining lexical matching with vector similarity so exact code matches are retrieved.
2. B Increase the similarity threshold so only strong semantic matches are returned.
3. C Reduce the chunk size so each chunk contains fewer competing terms.
4. D Re-embed the corpus more frequently so the codes become familiar.

14. The most relevant passage is retrieved but consistently ranks below less useful candidates. Which solution meets these requirements?

1. A Increase the embedding dimension used by the retriever.
2. B Reduce the candidate count so only the highest scoring passage is used.
3. C Increase the model's temperature so it weighs the context differently.
4. D Add a reranking model that scores each candidate jointly against the query before the top passages enter the prompt.

15. Users ask broad questions that require information from several distinct areas of the corpus. Which solution meets these requirements?

1. A Increase the chunk size so each chunk covers more of the corpus.
2. B Instruct the model to ask the user to narrow the question.
3. C Retrieve a larger number of candidates for the original query.
4. D Decompose the query into sub-queries, retrieve for each, and combine the results before generation.

16. Prompts used across several applications must be centrally versioned with an approval step before a change reaches production. Which solution meets these requirements?

1. A Store the prompts in each application's repository and review changes in pull requests.
2. B Store the prompts in environment variables set per environment.
3. C Store the prompts in a shared Amazon S3 object that each application reads at startup.
4. D Manage the prompts in Amazon Bedrock Prompt Management with versioned templates, and gate publication behind an approval workflow.

17. A complex task requires several model calls whose sequence depends on intermediate results. Which solution meets these requirements?

1. A Compose the calls as an Amazon Bedrock prompt flow with conditional branching between steps.
2. B Increase the model's maximum output tokens so it can complete all steps in one response.
3. C Combine every step into one long prompt and ask the model to perform them in order.
4. D Invoke the model repeatedly from application code with the sequence hardcoded.

18. A prompt change must be verified against known edge cases before it is published. Which solution meets these requirements?

1. A Publish the prompt to a single application first and observe user feedback.
2. B Review the prompt wording with a second engineer.
3. C Run the candidate prompt against a stored set of edge case inputs and compare the outputs against expected behaviour.
4. D Publish the prompt and watch the error rate for a day.

19. A proof of concept must establish whether a GenAI approach is viable before a team commits to building it. Which scope is appropriate?

1. A The narrowest path that exercises the riskiest assumption, measured against the acceptance criteria.
2. B The features the team finds most interesting to build.
3. C A complete implementation of every planned feature at reduced scale.
4. D The features that are quickest to implement.

20. An architect must document a GenAI design so other teams can reuse its decisions. Which content is most valuable?

1. A The names of the engineers who built it.
2. B A list of the AWS services the design uses.
3. C The constraints that drove each decision and the alternatives that were rejected.
4. D The estimated monthly cost of the design.

21. A model selection must be defensible to a reviewer who was not involved in the evaluation. Which evidence should be produced?

1. A The team's collective judgement after trying each candidate informally.
2. B The candidate with the largest context window and its specification.
3. C Measured results for each candidate on the same representative dataset with the criteria stated in advance.
4. D The vendor's published benchmark results for each candidate.

22. An application must degrade gracefully when its primary model is unavailable rather than failing outright. Which approach is appropriate?

1. A Retry the primary model until it responds.
2. B Return an error and let the calling application decide.
3. C Queue the request until the primary model recovers.
4. D Define a fallback behaviour, such as a smaller model or a cached response, and surface reduced capability to the user.

23. A fine-tuned model must be retired and replaced without interrupting the application. Which approach is appropriate?

1. A Deploy the replacement and switch all traffic at once during a maintenance window.
2. B Keep both versions serving indefinitely and let callers choose.
3. C Delete the current version and deploy the replacement immediately.
4. D Deploy the replacement alongside the current version, shift traffic gradually, and retire the old version once the new one is validated.

24. Input documents vary in quality, and some produce noticeably worse model responses than others. Which step should be taken first?

1. A Increase the model's context window so more of each document fits.
2. B Switch to a larger foundation model.
3. C Instruct the model to compensate for poor input quality.
4. D Profile the inputs that produce poor responses to identify what distinguishes them.

25. A conversation-based application must format its message history for the model correctly. Which consideration applies?

1. A Each turn must carry its role so the model distinguishes user input from its own previous responses.
2. B The history must be truncated to the most recent turn.
3. C The history must be summarized before every request.
4. D The history must be concatenated into a single block of text.

26. An application must extract entities from free text before passing structured values to a downstream system. Which approach is appropriate?

1. A Extract the entities with a purpose-built service or a schema-constrained model call, and validate the result against the expected structure.
2. B Pass the free text to the downstream system and let it parse the values.
3. C Store the free text and extract the entities in a nightly batch.
4. D Ask the model to return the entities and trust the format it produces.

27. A vector store must serve several tenants whose data must never be returned to another tenant. Which approach provides the strongest separation?

1. A Include a tenant identifier in the query text so the model filters appropriately.
2. B Separate indexes per tenant, or a tenant filter applied from a verified identity rather than from request content.
3. C Relying on embedding similarity to keep tenants' content apart.
4. D Add a post-processing step that removes other tenants' content from the response.

28. A metadata framework must support retrieval filtered by recency. Which attribute should be attached to each chunk?

1. A The source document's effective date, so a filter can exclude superseded content.
2. B The size of the source document in bytes.
3. C The number of chunks the source document produced.
4. D The date the chunk was embedded.

29. A vector index must be maintained as source systems publish changes continuously throughout the day. Which approach is appropriate?

1. A Rebuild the index nightly from the current source state.
2. B Reingest the entire corpus whenever any document changes.
3. C Poll the source system hourly and reingest everything that changed.
4. D Subscribe to change events from the source and upsert or delete the affected chunks as they arrive.

30. A chunking strategy must be chosen for a corpus of short, self-contained frequently asked questions. Which approach is appropriate?

1. A Split each answer across several chunks to improve precision.
2. B Combine several pairs into each chunk to reduce the index size.
3. C Treat each question and answer pair as a single chunk, since it is already a complete unit.
4. D Chunk at a fixed token count regardless of the pair boundaries.

31. An embedding model must be selected for a corpus in several languages. Which criterion is decisive?

1. A Whether the model was released most recently.
2. B Whether the model has the largest available embedding dimension.
3. C Whether the model is the same family as the generation model.
4. D Whether the model represents all the required languages in a shared space so cross-lingual retrieval works.

32. A query expansion step must improve retrieval for terse user questions. Which approach is appropriate?

1. A Lower the similarity threshold so weaker matches are included.
2. B Ask the user to rephrase the question more fully.
3. C Increase the number of candidates retrieved for the original question.
4. D Generate additional phrasings of the question and retrieve for each, combining the results.

33. A retrieval interface must be consumable by a foundation model as a callable tool. Which approach is appropriate?

1. A Include the retrieved passages in the system prompt before each request.
2. B Have the model generate a search query as free text for the application to interpret.
3. C Return the raw vector index contents to the model.
4. D Define the retrieval operation with a schema describing its parameters and return structure, and expose it as a function the model can call.

34. An application must maintain conversational context across turns while keeping the prompt within the context window. Which approach is appropriate?

1. A Retain only the most recent turn.
2. B Retain every turn verbatim until the window is exceeded, then start a new conversation.
3. C Retain recent turns verbatim and summarize older ones, preserving any facts the conversation depends on.
4. D Summarize every turn including the most recent.

35. A prompt must reliably produce output a downstream system can parse. Which approach is most reliable?

1. A Parse the response with a tolerant parser that accepts several formats.
2. B Instruct the model in the prompt to respond only in the required format.
3. C Lower the temperature so the format varies less.
4. D Constrain the output to a defined schema and validate it, rejecting or retrying responses that do not conform.

36. A prompt must be improved iteratively, and the team must know whether each change helped. Which approach is appropriate?

1. A Change the prompt and judge the improvement from a few sample responses.
2. B Adopt the version the team prefers on reading it.
3. C Maintain a fixed evaluation set and score each prompt version against it, keeping the version with the best result.
4. D Change several aspects of the prompt at once to make progress faster.

37. A prompt template must accept variable content without allowing that content to alter the instructions. Which approach is appropriate?

1. A Concatenate the variable content directly after the instructions.
2. B Place variable content in a clearly delimited section and instruct the model to treat it as data rather than instruction.
3. C Escape special characters in the variable content.
4. D Place the instructions after the variable content so they are read last.

38. A GenAI feature's business value must be validated before the team invests in production hardening. Which proof of concept design is appropriate?

1. A Demonstrate the feature to executives with hand-picked examples.
2. B Build production-grade infrastructure first so the proof of concept is realistic.
3. C Measure the feature's effect on the target business metric with a small user group before building resilience, monitoring, or scale.
4. D Compare the feature's output quality against a competitor's product.

39. A team is choosing between a synchronous request-response integration and an event-driven one for a GenAI feature. Which requirement favours event-driven?

1. A The user must see the response before proceeding.
2. B The feature's output is not needed immediately and generation may take longer than a user would wait.
3. C The feature is invoked once per user session.
4. D The feature must return within one second.

40. A model must be evaluated against capability limitations relevant to the use case before selection. Which approach is appropriate?

1. A Select the model with the most parameters.
2. B Test the candidate on the specific failure modes the use case is sensitive to, such as long context handling or structured output, rather than general benchmarks alone.
3. C Select the model the team is most familiar with.
4. D Select the model with the highest general benchmark score.

41. An application must switch to a backup model automatically when the primary returns errors above a threshold. Which implementation is appropriate?

1. A Schedule a job that checks the primary's health hourly.
2. B Retry the primary in a loop until it succeeds.
3. C Implement a circuit breaker in the invocation layer.
4. D Have operators flip a manual switch during an incident.

42. A fine-tuned adapter must be deployed to serve alongside the base model without duplicating the base weights. Which approach is appropriate?

1. A Merge the adapter into a full copy of the base weights and deploy the copy.
2. B Deploy a separate full model instance per adapter.
3. C Deploy the adapter as a separate lightweight artifact loaded onto the shared base model at inference.
4. D Retrain the base model with the adapter's data.

43. A data validation workflow must reject records that would exceed the model's input limits before invocation. Which check is required?

1. A Check the token count against the model's context limit, since character or byte length does not map directly to tokens.
2. B Check the byte length against a fixed limit.
3. C Check that the record is valid JSON.
4. D Check that the record is non-empty.

44. Audio recordings must be prepared for a text-based foundation model. Which processing step is required?

1. A Convert the audio to a different file format.
2. B Compress the audio to reduce its size.
3. C Transcribe the audio to text with a speech recognition service before passing it to the model.
4. D Encode the audio as base64 and include it in the prompt.

45. A request to a foundation model API is rejected with a validation error about the message structure. Which cause should be investigated first?

1. A Whether the API key is valid.
2. B Whether the model is available in the Region.
3. C Whether the request exceeds the rate limit.
4. D Whether the request body matches the model's expected schema, since different models on the same platform accept different formats.

46. A vector index must be sharded to handle a corpus too large for a single node. Which sharding consideration applies?

1. A Sharding removes the need for approximate search.
2. B Queries must fan out across shards and merge results, so shard count trades parallelism against merge overhead.
3. C Sharding is unnecessary if the embedding dimension is reduced.
4. D Each shard must hold a complete copy of the corpus.

47. A knowledge base must integrate documents from an internal wiki that changes frequently. Which approach is appropriate?

1. A Ask wiki editors to email updated pages.
2. B Use a connector that syncs from the wiki on a schedule or on change events, so the index tracks the source.
3. C Export the wiki manually once and ingest the export.
4. D Copy each wiki page into a document by hand.

48. A vector store's metadata must support filtering retrieval to documents a user is authorised to see. Which metadata design is appropriate?

1. A Attach the document's file size.
2. B Rely on the model to exclude unauthorised content.
3. C Attach an access-scope attribute to each chunk derived from the source document's permissions, and filter on it from the user's verified entitlements.
4. D Attach the document's author name only.

49. Retrieval returns the correct document but the chunk lacks the surrounding context the model needs. Which approach is appropriate?

1. A Reduce the chunk size further.
2. B Increase the number of unrelated chunks retrieved.
3. C Retrieve the matched chunk together with its neighbouring chunks, or use hierarchical chunking that returns the parent section.
4. D Instruct the model to infer the missing context.

50. A retrieval system must handle a query that is really two separate questions. Which approach is appropriate?

1. A Answer only the first question.
2. B Retrieve for the combined query and hope both are covered.
3. C Ask the user to submit one question at a time.
4. D Decompose the query into its sub-questions, retrieve for each, and answer each part.

51. A vector search solution must be deployed with the least operational burden for a modest corpus. Which option is appropriate?

1. A A Bedrock Knowledge Base with its managed vector store, which handles ingestion, embedding, and indexing.
2. B A custom index built in application memory.
3. C A self-hosted vector database on EC2.
4. D A self-managed OpenSearch cluster.

52. An embedding model must be evaluated for a domain before adoption. Which evaluation is appropriate?

1. A Check whether the model is the most recently released.
2. B Compare the model's embedding dimension against alternatives.
3. C Measure retrieval recall against a labelled set of domain queries and relevant passages.
4. D Measure the model's embedding generation speed.

53. A conversational application must recognise the user's intent to route to the right handling flow. Which approach is appropriate?

1. A Route every message to the same handler.
2. B Send every message to the largest model and let it decide.
3. C Classify the intent with a lightweight model or service before invoking the main flow, so routing is deterministic and cheap.
4. D Ask the user to select their intent from a menu.

54. A prompt must include worked examples to guide the model's output format. Which consideration applies?

1. A Examples consume context tokens on every request, so their number should be the minimum that achieves the format reliably.
2. B Examples should be placed after the user's input.
3. C Examples should be different on every request.
4. D More examples are always better.

55. A chain-of-thought instruction improves reasoning quality but exposes the reasoning to the user. Which approach is appropriate?

1. A Remove the chain-of-thought instruction.
2. B Reduce the temperature so reasoning is shorter.
3. C Show the reasoning to the user.
4. D Instruct the model to reason in a delimited section and strip it before returning the response.

56. A prompt regression test must confirm that a change did not degrade output on previously correct cases. Which test design is appropriate?

1. A A comparison against the model provider's documentation.
2. B A manual review of a few outputs after each change.
3. C A fixed set of inputs with recorded expected outputs, run after every prompt change with a comparison against the baseline.
4. D A test only when a user reports a problem.

57. A GenAI solution's requirements must be captured before design begins. Which information is essential?

1. A The task the system performs, the quality bar it must meet, and the constraints on latency, cost, and data handling.
2. B The foundation model the team intends to use.
3. C The cloud services the team is familiar with.
4. D The programming language the team prefers.

58. An architecture decision for a GenAI application must be recorded so it can be revisited. Which record is most useful?

1. A A diagram of the resulting architecture.
2. B The decision, the constraints that drove it, the alternatives considered, and the conditions under which it should be revisited.
3. C The name of the person who decided.
4. D The date the decision was made.

59. A model must be selected for a task requiring reasoning over long documents. Which approach is appropriate?

1. A Evaluate candidates on long-document tasks representative of the workload, measuring both accuracy and cost at that context length.
2. B Select the model with the largest context window.
3. C Select the model with the most parameters.
4. D Select the model with the highest general benchmark score.

60. A foundation model's inference parameters must be configured for a task requiring consistent structured output. Which approach is appropriate?

1. A Set a low temperature and constrain the output format, then validate the result against the expected structure.
2. B Increase the maximum output tokens.
3. C Set a high temperature to increase variety.
4. D Increase the top-p value.

61. An application must be able to switch foundation models without rewriting its integration. Which approach is appropriate?

1. A Fine-tune every candidate model on the same data.
2. B Abstract the model invocation behind an internal interface so the provider-specific details are isolated.
3. C Use only models from a single provider.
4. D Call each provider's API directly from the application logic.

62. Documents of mixed quality must be prepared for a knowledge base. Which approach is appropriate?

1. A Ingest every document regardless of quality.
2. B Ingest only documents already in plain text.
3. C Convert every document to images before ingestion.
4. D Assess each document's extractability, normalize the usable ones, and exclude those that cannot be processed reliably.

63. Personally identifiable information must be removed from data before it reaches a foundation model. Which approach is appropriate?

1. A Detect and redact the information in a preprocessing step before the content enters the prompt or the index.
2. B Filter the model's response for personal information.
3. C Instruct the model not to repeat personal information.
4. D Encrypt the data before sending it to the model.

64. Conversation history must be formatted so a foundation model distinguishes system instructions from user input. Which approach is appropriate?

1. A Use the model's message role structure so each part carries its role explicitly.
2. B Place the instructions at the end of the history.
3. C Prefix each line with a label in plain text.
4. D Concatenate everything into a single text block.

65. A vector store must be selected for a corpus that will grow to hundreds of millions of vectors. Which approach is appropriate?

1. A Choose a store that supports horizontal scaling and approximate nearest neighbour search at that scale.
2. B Choose the store with the simplest API.
3. C Choose the store the team used for a smaller corpus.
4. D Choose an in-memory store for lowest latency.

66. Vector index parameters must be tuned for a workload that prioritizes recall over latency. Which approach is appropriate?

1. A Increase the number of index shards.
2. B Reduce the embedding dimension.
3. C Reduce the number of vectors indexed.
4. D Adjust the index's search parameters toward higher recall, accepting the additional query time.

67. A vector store must support filtering by document attributes at query time. Which approach is appropriate?

1. A Create a separate index per attribute value.
2. B Store the attributes in a separate database and filter afterwards.
3. C Store the attributes as metadata alongside each vector and apply the filter during retrieval.
4. D Encode the attributes into the embedding.

68. A vector index must be rebuilt without interrupting queries. Which approach is appropriate?

1. A Build the new index alongside the current one and switch queries to it once it is ready.
2. B Rebuild the index in place while queries continue.
3. C Delete the current index and rebuild it.
4. D Pause queries during the rebuild.

69. Embeddings must be regenerated after the embedding model is changed. Which approach is appropriate?

1. A Re-embed only documents added since the change.
2. B Keep the existing vectors and add new ones with the new model.
3. C Convert the existing vectors to the new model's space.
4. D Re-embed the entire corpus with the new model, since vectors from different models are not comparable.

70. A vector store's cost must be controlled for a corpus where most documents are rarely retrieved. Which approach is appropriate?

1. A Index only the most recent documents.
2. B Reduce the embedding dimension for all documents.
3. C Delete documents that have not been retrieved.
4. D Tier the corpus so rarely retrieved content is held in cheaper storage and promoted when access patterns change.

71. Retrieval must return relevant results for queries using terminology absent from the corpus. Which approach is appropriate?

1. A Lower the similarity threshold.
2. B Combine semantic retrieval with query expansion or a synonym mapping so vocabulary differences are bridged.
3. C Increase the number of results retrieved.
4. D Instruct the model to rephrase the answer.

72. A reranking step must improve the ordering of retrieved passages. Which approach is appropriate?

1. A Return the candidates in retrieval order.
2. B Reorder the candidates by length.
3. C Score the retrieved candidates with a model that evaluates query and passage together, then reorder by that score.
4. D Reorder the candidates by document date.

73. Retrieval must weight recent documents more highly without excluding older ones. Which approach is appropriate?

1. A Combine the similarity score with a recency factor so newer documents rank higher while older ones remain eligible.
2. B Filter out documents older than a threshold.
3. C Retrieve only the most recent documents.
4. D Sort the results purely by date.

74. A chunking strategy must be chosen for long technical documents with nested sections. Which approach is appropriate?

1. A Chunk at a fixed token count regardless of structure.
2. B Chunk on section boundaries so each chunk is a coherent unit, with parent context available for retrieval.
3. C Treat each document as a single chunk.
4. D Chunk on sentence boundaries.

75. Retrieval quality must be measured before the application is deployed. Which approach is appropriate?

1. A Build a labelled set of queries with known relevant passages and measure recall and ranking against it.
2. B Review a sample of retrieved results informally.
3. C Measure the retrieval latency.
4. D Count the number of documents indexed.

76. A retrieval system must handle a query for which no relevant content exists. Which approach is appropriate?

1. A Return the closest matches regardless of score.
2. B Detect that no candidate exceeds the relevance threshold and respond that the information is not available rather than generating from weak matches.
3. C Lower the threshold until results are returned.
4. D Generate an answer without retrieval.

77. Retrieval must span several knowledge bases with different access permissions. Which approach is appropriate?

1. A Query the bases the application has access to.
2. B Query only the bases the user is entitled to, determined from their verified identity before retrieval.
3. C Merge the bases into one index.
4. D Query all bases and filter the results by permission.

78. A system prompt must define behaviour that user input cannot override. Which approach is appropriate?

1. A Place the behaviour in the system role and treat user content as data, while limiting what the application can do so an override has bounded impact.
2. B Place the behaviour at the end of the user message.
3. C Repeat the behaviour several times in the prompt.
4. D Instruct the model to ignore contradicting instructions.

79. A prompt must produce output in a specific language regardless of the input language. Which approach is appropriate?

1. A Translate the input before sending it.
2. B Translate the response after generation.
3. C State the required output language explicitly in the instructions and validate the response language.
4. D Assume the model matches the input language.

80. A prompt library must be maintained as the application grows. Which approach is appropriate?

1. A Store prompts in a shared document.
2. B Keep prompts inline in the application code.
3. C Keep each developer's prompts locally.
4. D Store prompts as versioned artifacts with their evaluation results, so changes are traceable and reversible.

81. A prompt's instructions conflict with one another in edge cases. Which approach is appropriate?

1. A Add more instructions to cover each case.
2. B Increase the model's temperature so it varies its choice.
3. C Order the instructions by precedence and state how conflicts should be resolved.
4. D Remove the conflicting instructions.

82. A prompt must adapt its examples to the type of request received. Which approach is appropriate?

1. A Use the same examples for all requests.
2. B Select the example set at run time from the request's classification, so the examples match the task.
3. C Omit examples entirely.
4. D Include every example in every prompt.

83. A prompt must be prevented from leaking its system instructions to users. Which approach is appropriate?

1. A Filter the response for content resembling the system instructions and limit what the instructions reveal in the first place.
2. B Encrypt the system prompt.
3. C Instruct the model not to reveal its instructions.
4. D Shorten the system prompt.

84. A prompt's token cost must be reduced without degrading output quality. Which approach is appropriate?

1. A Truncate the prompt to a fixed length.
2. B Remove all examples from the prompt.
3. C Compress the prompt text.
4. D Measure which parts of the prompt affect the output and remove those that do not.