AWS Generative AI Developer practice

{{ summaryValue }}

{{ summaryLine }}

Build an AIP-C01 practice round

Unofficial AWS Certified Generative AI Developer – Professional (AIP-C01) practice · current online guide snapshot verified September 11, 2026. Independently authored; not affiliated with or endorsed by AWS.

40 original English production scenarios cover selected concepts in all five domains and 20 tasks. This is untimed single-choice practice with uniform sampling. The real exam has 75 questions in 180 minutes, including multiple response; this bank does not simulate its full format or difficulty.

Official guide and format · Official preparation resources

Blueprint weights: models/data 31%; integration 26%; safety/governance 20%; operations 12%; evaluation 11%. These describe official scope, not this round’s sampling weights. Applied and multiple-constraint scenarios are author-assigned complexity levels. No prior certification is required; AWS recommends production development and hands-on GenAI experience.

Optional. Blank uses a stable default; a custom seed changes question and option order. Long seeds use a whole-character prefix within the 128 UTF-16-unit limit. Invalid Unicode uses the default. Active and replay seeds do not overwrite this optional entry.

{{ seedNotice }}

{{ examSetupValidation }}

{{ answeredCount }} of {{ questionCount }} answered · {{ Math.round(progressPercent) }}%

{{ progressStatusLine }}

Question {{ examCursor + 1 }} of {{ questionCount }}

{{ currentQuestion.stem }}

{{ currentQuestion.accessible_clue }}

Answer recorded. Corrections appear when you view results.

Review this practice round

{{ resultLead }}

AIP-C01 · online guide snapshot September 11, 2026 · original bank 2026.09.11. This score describes only the items answered. It is not an AWS scaled score, a pass prediction, or evidence of certification. AWS and its service names are trademarks of Amazon.com, Inc. or its affiliates.

Round seed: {{ examSession.seed }}. Continue with hands-on AWS work and the full official guide; a perfect round covers only this curated selection.

{{ readinessReviewTitle }}

{{ readinessReviewLine }}

{{ card.label }}
{{ card.value }}
{{ card.note }}

Missed answer review

  1. {{ item.prompt }}

    Your answer: {{ item.selected_answer_label }}

    Correct answer: {{ item.correct_answer_label }}

    {{ item.explanation }}

    {{ item.source.label.split(' — ')[0] }} · {{ item.source.version }}

Domain performance

Correct and missed answers by official domain for this round. Domains absent from the chart were not sampled. Use the missed-item sources to review that domain, then practise a focused set and check the full official guide.

Loading domain-performance chart…

{{ examChartStatus }}

{{ examChartExportStatus }}

Answer ledger

Attempt / question ID Prompt Topic Your answer Correct answer Outcome Explanation Source Copy
{{ `AIP-C01 | guide snapshot 2026-09-11 | bank ${examSession.bank_version} | seed ${JSON.stringify(examSession.seed).replace(/\s/g, character => '\\u' + character.charCodeAt(0).toString(16).padStart(4, '0'))} (JSON) | ${row.question_id}` }}{{ ` | ${examSession.session_id} | ${examSession.bank_id} | session version ${examSession.session_version} | ${examSession.set_id} | ${examSession.mode_id} | ${examSession.count} questions` }} {{ row.prompt }} {{ row.category_label }} [{{ row.category_id }}] {{ row.selected_answer_label }} [{{ row.selected_answer_id }}] {{ row.correct_answer_label }} [{{ row.correct_answer_id }}] {{ row.outcome_label }} {{ row.explanation }} {{ row.source ? `${row.source.label.split(' — ')[0]} · ${row.source.version}` : 'Official guide' }}{{ sourceUrl(row.question_id) }}

{{ examLedgerExportStatus }}

AWS AI Practitioner Exam Prep Practise AIF-C01 with 40 independently authored questions and sourced explanations, then review missed answers across five exam domains.
AWS Developer Associate Exam Prep Practise AWS Developer Associate DVA-C02 with 28 original questions, source-linked explanations and focused retries of missed answers.
Large Language Model (LLM) Usage Cost Calculator Estimate monthly LLM API cost from token use, caching and retries with budget checks plus same-workload model price comparisons.
IAM Policy Permissions Analyzer Review AWS IAM policy JSON for wildcard breadth, privilege paths, and policy-shape issues before validating effective access in AWS.

An assistant can quote an obsolete maintenance procedure in fluent, confident language. A reservation agent can identify the intended cancellation but still need permission and customer approval before performing it. Production generative artificial intelligence (AI) development requires the evidence, model response and permitted actions to work together within quality, response-time and data-handling constraints.

The AWS Certified Generative AI Developer – Professional (AIP-C01) exam concerns this application-development work on Amazon Web Services (AWS). A foundation model is a broadly trained model used as the basis for tasks such as summarization, document questions and information extraction. Amazon Bedrock supplies model access and features for connecting models to application data and workflows. Production integration involves choosing suitable models, supplying useful context, controlling access, evaluating responses and diagnosing failures. Developing and training a foundation model from scratch is outside the exam’s stated target role.

For a document assistant, retrieval-augmented generation (RAG) supplies relevant source passages to a model before it produces an answer. The source collection must be indexed and kept current, and retrieval must find the passages that address the request. More elaborate instructions cannot refresh an outdated index, and a larger model does not establish that the application retrieved the right evidence.

Grounding
Whether the response is supported by the reference material supplied to it.
Relevance
Whether the response addresses the user’s question.

These properties can disagree. An answer may accurately describe one maintenance procedure while the user asked about another. An unsupported claim may also be directly relevant to the question. Checking both properties is more informative than judging fluency alone, but even a grounded answer can inherit an error from its source.

Agents introduce a further responsibility because they can request operations through connected software. A proposed cancellation is not evidence that a cancellation occurred, and instructions found in a retrieved document are not the customer’s consent. The application still needs to authenticate the user, authorize the operation and validate its arguments. Content filters and isolated sessions address particular risks; neither replaces those checks.

Quality must also survive changes in prompts, models and source data. Representative evaluation examples should include long documents, missing evidence and relevant groups of users, rather than only polished demonstrations. A smaller model or shorter context may reduce processing cost, but it is suitable only where testing preserves the required answer quality. Measurements of retrieval time, time to the first response token and total generation time help identify what actually became slower.

The recommended background is at least two years building production-grade applications on AWS or with open-source technologies, general AI, machine-learning or data-engineering experience, and one year implementing generative AI solutions. No prior certification is required. Scenario practice supports study, but selecting an answer does not demonstrate that someone can deploy, secure or operate the corresponding system.

How to Use This Tool:

This is independently authored, unofficial practice with 40 original English scenarios, not AWS examination questions. It is not affiliated with or endorsed by AWS. A new setup starts with five questions from All domains and feedback after each answer.

  1. Choose a Domain / practice focus. Use the full pool, one of the five domains, or an author-assigned group such as Multiple-constraint scenarios. The two complexity groups are author-assigned, not calibrated levels of official exam difficulty.
  2. Choose Feedback timing. Learn — feedback after each answer reveals the correction, explanation and source after each choice. Review — feedback at completion records the choice but withholds corrections until results.
  3. Select a Question count. Choose an offered length, up to 40 for the full pool; focused menus reflect their smaller pools. Changing focus returns an unsupported count to five.
  4. Start with Start exam practice. Before starting, an optional Replay seed under Advanced can identify a repeatable question sequence. Leaving it blank uses a stable default, not a fresh random draw.
    If Start exam practice is unavailable, choose an available focus, feedback mode and question count.
  5. Select one of the four answers, then continue with Next question. Selection immediately locks the answer in both modes; there is no separate submission step. After the last answer, choose View results.
    End round clears the active attempt and returns to setup. It does not grade an unfinished round.
  6. Review Missed answer review and the explanations in the Answer ledger. Choose Retry missed for just the incorrect items, Replay same seed to repeat the round, or Try a new seed for another seeded selection. Change setup returns to the focus and count choices.

Interpreting Results:

During a round, the headline counts answers submitted, not answers correct. A progress bar at 100% means every question has an answer. The completed headline instead reports correct answers out of the round’s total.

Domain performance shows counts of correct and missed answers, not weighted domain scores. A longer correct bar can reflect more sampled questions rather than stronger knowledge. A domain absent from the chart was not sampled and is untested, not passed or failed.

The Answer ledger connects each scenario to your choice, the authored correct choice, its explanation and its source. Review the requirement that made an alternative unsuitable, such as stale indexed data, missing authorization or unsupported model capabilities. This is more useful than memorizing an answer’s screen position, which can change between seeded rounds.

Perfect round means no mistakes on the selected items. It is not a readiness band, a pass prediction, a certificate or evidence of professional competence. Replays also involve familiar material. Use the missed-item sources to guide hands-on study and compare the selected topics with the full official guide before drawing broader conclusions.

Technical Details:

Each scenario has one authored best answer for its stated requirements. Agreement with that answer can be counted exactly, but the count does not measure deployment skill or account for differences in official exam difficulty.

Counting answers, accuracy and completion

A correct answer earns one point and an incorrect answer earns zero. There is no partial credit, negative marking or domain multiplier. The score, accuracy and progress use the following rules.

Score points = C Accuracy (%) = 100×CN Progress (%) = 100×AN

Here, C is the number of correct answers, A is the number answered and N is the total number of questions in the round. All are whole-number counts, with at least one question in a valid round. Possible points equal N, and the score ranges from zero to N.

Unanswered questions remain in the accuracy denominator but do not become incorrect answers. Misses are answered questions that were wrong. Accuracy keeps the unrounded ratio. The visible progress percentage rounds to the nearest whole percent, with an exact half-percent tie rounding upward. The completed correct-count fraction remains exact.

Condition Result meaning
Fewer than all questions answered The round is incomplete. Progress is available, but the completed review is not ready.
All questions answered, with at least one incorrect Start with the missed answers identifies items for review.
All questions answered correctly Perfect round indicates zero misses, without creating an official pass result.

Official scope and the smaller practice bank

The questions use an AIP-C01 guide snapshot from September 11, 2026 and bank version 2026.09.11. That is a study-material date, not the exam’s launch date. The bank’s stated reach across all 20 guide tasks means selected concepts within those tasks, not complete testing of every task or skill. It is not an exhaustive or difficulty-calibrated exam simulation.

Official domain percentages describe the examination blueprint. The practice counts below describe the authored question pool and do not impose those percentages on a round.

Official domain Official weight Practice items Selected practice concepts
Foundation Model Integration, Data Management, and Compliance 31% 12 Model evaluation, ingestion, embeddings, retrieval, prompt versions and context budgets.
Implementation and Integration 26% 10 Agent actions, session ownership, memory, message interfaces, retries and workflows.
AI Safety, Security, and Governance 20% 8 Grounding, sensitive data, private connectivity, authorization, source history and fairness.
Operational Efficiency and Optimization for GenAI Applications 12% 5 Prompt caching, batch inference, latency, operation tracing and context reduction.
Testing, Validation, and Troubleshooting 11% 5 Retrieval evaluation, model judges, access failures, agent traces and regression testing.

The official exam has 75 questions in 180 minutes, including multiple-choice and multiple-response questions. It includes 65 scored questions and 10 unidentified unscored questions. Multiple-response items require all correct choices for credit. This practice is untimed, offers one answer from four choices and scores every selected item.

Official results use a scaled score from 100 to 1,000, with 750 as the minimum passing score. That does not make 75% correct on this bank an official passing result. There is no published conversion from this practice percentage to the AWS scale.

Repeatable question selection

Selection is uniform within the chosen pool and without replacement within a round. It does not enforce official domain quotas or adapt question difficulty to earlier answers. Taking all 40 questions uses the entire bank; a shorter mixed round can omit a domain.

Keeping the same seed, focus, question count and feedback timing repeats the questions and their order while the bank remains unchanged. A new seed can change the selection or ordering, but different rounds can overlap. It cannot create additional authored questions, and selecting an entire pool necessarily repeats its membership.

Surrounding spaces in a seed are ignored. Long text is shortened at a whole character: the allowance holds 128 basic letters or digits, while many emoji take two places. Blank or unusable text falls back to the stable default. Replaying does not overwrite your optional entry, so use the Round seed in results to identify the text actually used.

Retrieval failures need different repairs

An embedding model turns text into a numeric vector. A vector database compares compatible query and document vectors to find candidates; the embedding model does not search the document collection by itself. Ingestion, candidate selection and candidate ordering can therefore fail independently.

Observation Mechanism or limit Appropriate check
An old procedure is still retrieved Replacing a source file does not itself refresh the existing index. Sync the data source, check ingestion completion and verify the retrieved revision.
Related passages appear, but an exact error code is missed Semantic-only retrieval lacks the additional text-search signal of hybrid retrieval. Evaluate hybrid retrieval on a supported store with a filterable text field.
The relevant passage is present but ranked poorly Reranking changes the order of retrieved candidates; it cannot recover absent candidates. Compare ranking quality and added latency with a supported reranker.
Hierarchical retrieval returns fewer entries than requested The requested count is a maximum. Child matches can be consolidated into shared parent chunks. Check parent substitution before concluding that documents were lost.

Retrieve-only evaluation examines the evidence returned before response generation is considered. Supported evaluation models and configurations vary by AWS Region. Grounding and relevance then assess different properties of the answer. Bedrock contextual grounding checks support document question answering, summarization and paraphrasing. The documented feature does not support conversational question-answering or chatbot use cases, and its checks do not establish universal factual truth.

Context budgets and operating costs

When a scenario declares one shared budget for input and output tokens, the response reserve must fit within that budget and be deducted before admitting input. Tokens are the model’s counted units of content, not a fixed number of characters or words.

Imax = B − R

B is the declared total context budget, R is the output reserve and Imax is the maximum combined input, all measured in tokens. Instructions, conversation, retrieved passages and tool descriptions must fit together. This subtraction describes the stated shared-budget scenario, not a universal limit for every Bedrock model; actual counting and limits remain model-specific.

Converse provides a common message interface for supported models, but does not make all model capabilities interchangeable. ConverseStream enables incremental delivery when the application handles and forwards the response events. Earlier visible output does not guarantee a shorter total generation time.

For a model supporting explicit prompt caching, reusable static content should precede changing content and use the supported cache controls. Reusing processed prompt content is different from returning a cached answer. It does not eliminate response generation or guarantee identical answers. Likewise, reducing irrelevant retrieved text is justified only while measured answer quality remains acceptable.

Agent actions and data boundaries

Before a consequential action such as cancellation, configured user confirmation supplies an approval step. Authorization and argument validation remain necessary even after approval. AgentCore isolates sessions but does not enforce which user owns each session. The application must bind session access to the authenticated user.

Bedrock sensitive-information filters evaluate text, not the contents of tool-call arguments, tool results or tool definitions. Those fields require their own validation, redaction and access controls. A filtered conversation is not evidence that every downstream payload has been checked.

Invocation logging is disabled by default and is separate from aggregate request metrics. For supported Bedrock runtime calls, approved CloudWatch Logs or Amazon S3 destinations must be in the same account and Region. Payload logging and agent traces can expose sensitive content, so access and retention decisions belong in the diagnostic design.

Cross-Region inference profiles define possible processing destinations. Every possible destination in a geographic profile must be checked against the application’s geographic restrictions, and the selected model must support that profile. A global profile does not pin processing to the request’s source Region, and an inference profile is not a reservation of provisioned throughput.

Evaluation beyond a convincing demonstration

A large language model (LLM) judge provides computed ratings, not an infallible reference. Comparing its ratings with domain experts on representative held-out examples tests whether its judgments fit the task. Release checks should use versioned regression cases, including absent evidence and other edge cases, rather than only examples used to tune the prompt. Aggregate improvements can also hide worse results for a particular dialect or user group. Investigating those differences and retesting a mitigation supplies evidence that an overall average cannot provide.

Privacy Notes:

Practice scoring takes place in your browser. The saved page address contains the round selection, seed, submitted answers and current position without encryption. Someone receiving that address can restore those answers. Addresses can remain in browser history and reach the site when opened or reloaded, so this is not a never-leaves-device workflow.

A saved address is the restoration mechanism, not a separate study-history account. An incompatible or malformed saved attempt is rejected rather than silently treated as a valid result. Start a new round when Saved attempt unavailable appears. If saving fails, the earlier saved link is cleared; keep the page open and use Retry save or end the round.

Use a non-sensitive replay seed, and treat saved links and exported answer ledgers as records of your responses. Ending a round clears its active saved address, not earlier copies, bookmarks or browser history. Loading supporting chart or document features and opening source references can use the network, so fully offline operation is not promised.

Worked Examples:

One missed ingestion question in a full round

Choose All domains, 40 questions and feedback at completion. Suppose the only mistake is choosing a generation-setting change for the stale-procedure scenario. The result is 39/40, equivalent to 97.5%, and the models/data domain contains 11 correct answers and one miss. Retry missed creates a one-question round even though normal setup menus begin at two. Choosing the synchronization answer then produces 1/1 for that retry, not a revised 40/40 result for the original round.

Reserving space for the generated response

When the context-budget scenario appears, it declares 8,192 tokens in total and reserves 1,024 for the response. Subtracting the reserve leaves 7,168 tokens for all input components, so select the answer with that combined limit. Allocating 7,168 tokens to retrieved passages alone would leave no room for instructions, conversation or tool schemas. These figures belong to the question’s declared budget, not to every model offered through Bedrock.

References: