Study guide for the AWS Certified Generative AI Developer - Professional exam
AWS · October 2026 exam blueprint · 232 curriculum items
Start guide
0.1 Classify models by the required output
About this guide
Each bite isolates one concept or one architecture decision. Read them in order to prepare for the exam, or come back to a single bite later as a reference — the shape is the same either way, so you always know where in a lesson an answer will be.
Lessons are made to fit into the gaps between real engineering work, including the time while your coding agent is running. There is no reading-time promise attached to that, only a deliberate scope: one decision per bite.
A Lab completes when you open its official exercise — and, if it ends with a knowledge check, once you have answered that too. Anything else that carries a knowledge check — a bite, a checkpoint, a mock exam — completes when you answer every question of it; a bite without one completes when you say so. Completion only accumulates — nothing here takes it back.
Prerequisites
Outside the exam blueprint0 of 16 items completed
- 0.1 Classify models by the required outputDistinguish foundation models, language models, and multimodal models by their supported inputs, outputs, and application roles.
- 0.2 Budget context before discarding inputExplain model-specific token counts, distinguish context and output limits, and recognize when truncation changes the task.
- 0.3 Separate sampling controls from completion limitsConfigure supported generation parameters and distinguish candidate-token sampling from output length and stopping behavior.
- 0.4 Align vector representations before comparing similarityExplain embeddings, dimensions, and similarity metrics, and identify the representation contract required for semantic search.
- 0.5 Validate outputs against model limitationsRecognize capability, knowledge, variability, and output constraints, and choose checks that address each failure mechanism.
- 0.6 Choose delivery by the completion contractCompare complete responses, incremental streams, deferred invocations, and dataset jobs by how callers observe and accept results.
- 0.7 Choose instructions and examples deliberatelyChoose role instructions, zero-shot or few-shot examples, and prompt structure while separating requested behavior from enforced output shape.
- 0.8 Separate instructions from evidence and supportDistinguish prompt engineering, context engineering, and grounding to choose the intervention that addresses an answer's failure.
- 0.9 Locate failures across the RAG stagesExplain ingestion, retrieval, augmentation, and generation and identify which stage must change when a RAG answer fails.
- 0.10 Choose who controls the next actionExplain agent loops and tool calling and choose deterministic or agentic control from the workflow's ordering and decision requirements.
- 0.11 Assign state to the required lifetimeDistinguish agent sessions, runtime working state, and durable memory so continuity survives the intended lifecycle without crossing user boundaries.
- 0.12 Read quality and runtime evidence separatelyInterpret evaluation metrics, logs, traces, and response signals without confusing successful inference with acceptable answers.
- 0.13 Trace permissions from principal to requestIdentify the calling principal, separate role trust from permissions, and constrain GenAI workloads with temporary credentials and least privilege.
- 0.14 Assign infrastructure to each application boundaryDistinguish network paths, compute, object storage, work queues, and event routing in a GenAI application architecture.
- 0.15 Resolve model targets before invoking BedrockDiscover Bedrock model targets, verify account access, and distinguish inference profiles and runtime calls from control-plane operations.
- 0.16 Separate model resources from serving endpointsDistinguish SageMaker AI model resources, inference containers, endpoint configurations, deployed endpoints, and JumpStart model discovery.
Foundation Model Integration, Data Management, and Compliance
31% of the exam0 of 68 items completed
Analyze requirements and design GenAI solutions
- 1.1.1 Turn business goals into testable requirementsTranslate a GenAI use case into measurable behavior, quality gates, operational constraints, and safe fallback outcomes.
- 1.1.2 Choose the pattern your evidence requiresChoose direct inference, retrieval, agents, or model customization by identifying the missing capability and its trust boundary.
- 1.1.3 Match inference hosting to operating constraintsCompare managed foundation model APIs with hosted deployments using measured latency, demand, residency, and operational responsibilities.
- 1.1.4 Make your Bedrock experiment decide investmentDesign a measurable Amazon Bedrock proof of concept with representative tests, risk gates, unit economics, and an explicit investment decision.
- 1.1.5 Review GenAI designs with verifiable evidenceUse the AWS Generative AI Lens and Well-Architected Tool to review a complete workload, record risks, and track improvements.
- 1.1.6 Package approved architectures for repeatable deploymentStandardize GenAI infrastructure with reusable CDK definitions, CloudFormation deployment review, and governed Service Catalog products.
- LABDeploying a CI/CD Pipeline with a Managed Deployment Stage using AWS CDKBuild and test a containerized application pipeline, define its infrastructure with AWS CDK, and add deployment stages.
- LABDeploying and managing Infrastructure as CodeUpdate a deployed CloudFormation template with Application Composer, tag a Resource Group, and detect configuration drift.
Select and configure FMs
- 1.2.1 Eliminate models that cannot serve the taskBuild a foundation-model shortlist from required capabilities, context budgets, regional access, and documented operating limits.
- 1.2.2 Benchmark the complete model operating envelopeCompare eligible foundation models using representative quality gates, load measurements, token usage, and cost per accepted result.
- 1.2.3 Move model choices behind a validated contractSeparate model-specific interfaces from application behavior and release model routing through validated AWS AppConfig configuration.
- 1.2.4 Choose inference routes within processing boundariesCompare regional inference with geographic and global Bedrock inference profiles using capacity, authorization, and processing-location constraints.
- 1.2.5 Bound failover before adding fallback modelsDesign application model fallback with explicit error classes, shared attempt budgets, circuit breakers, and safe terminal outcomes.
- 1.2.6 Match LoRA serving to adapter demandChoose merged weights or SageMaker AI multi-adapter inference and preserve the base-model, adapter, and runtime compatibility contract.
- 1.2.7 Approve a reproducible model release packageUse SageMaker Model Registry versions, evaluation evidence, and approval states to govern custom model artifacts before deployment.
- 1.2.8 Promote releases with a tested return pathPromote approved model configurations across environments and choose SageMaker deployment guardrails or AppConfig rollback for the changed resource.
- 1.2.9 Replace retiring models before access disappearsTrack Bedrock model lifecycle by version and Region and validate replacements before retirement removes production or recovery access.
- 1.2.10 Fail over the layer that actually failedChoose Route 53 DNS failover for regional API recovery and application fallback for failures of an individual model path.
- 1.2.11 Replicate the assets recovery actually needsDesign S3 Cross-Region Replication around versioned model assets, encryption permissions, replication lag, and verified recovery readiness.
- LABConfigure DNS and Routing Policies with Amazon Route 53Configure Route 53 private DNS, endpoint records, and simple routing with health checks.
- LABOrchestrate ML Workflow using Amazon SageMaker Pipelines and Model RegistryBuild a SageMaker Pipeline that trains or evaluates a model and registers the result in Model Registry.
- LABCreate functions using AWS SDK for Python with Amazon Bedrock APIInvoke Amazon Bedrock models with Boto3, wrap a call in a parameterized function, and generate text and images with Amazon Titan.
Implement data validation and processing pipelines for FM consumption
- 1.3.1 Set quality gates before model consumptionDefine measurable input thresholds and route incomplete, inconsistent, invalid, or unsuitable data before foundation model consumption.
- 1.3.2 Separate input profiling from runtime healthCombine Glue quality rules, Data Wrangler insights, and Lambda metrics to distinguish defective inputs from processing failures.
- 1.3.3 Choose extraction by the required evidenceSelect AWS extraction and understanding services by input modality, required output structure, and evidence needed to validate the result.
- 1.3.4 Match preprocessing compute to work unitsChoose SageMaker Processing or Lambda from the transformation's runtime, resource needs, arrival pattern, and artifact release boundary.
- 1.3.5 Preserve evidence through multimodal preprocessingDesign modality-specific preprocessing branches that retain source provenance and release only validated, traceable artifacts.
- 1.3.6 Separate instructions from Converse message contentBuild Converse requests that preserve instruction ownership, message roles, content-block structure, and model capability constraints.
- 1.3.7 Validate native payloads before model invocationValidate InvokeModel request bodies against the selected model's required fields, parameter constraints, and native response contract.
- 1.3.8 Match endpoint bytes to container contractsPrepare SageMaker AI inference requests and decode responses using explicit MIME types, container schemas, and payload boundaries.
- 1.3.9 Preserve conversation meaning within context budgetsAssemble ordered conversation history, count model-specific input tokens, and reduce context without breaking message dependencies.
- 1.3.10 Validate enrichment against the extracted evidenceChoose Comprehend or Bedrock for enrichment and validate transformed records against schema, provenance, and business constraints.
- LABConfigure Data Pipelines with AWS GlueJoin S3 datasets in AWS Glue, orchestrate incremental processing, and validate the output with Athena.
- LABAnalyze and Prepare Data with Amazon SageMaker Data Wrangler and Amazon EMRProfile and prepare a dataset in SageMaker Data Wrangler, then process it with Amazon EMR.
Design and implement vector store solutions
- 1.4.1 Choose stores by retrieval and integrationCompare vector and graph stores by retrieval needs, operational boundaries, and Amazon Bedrock Knowledge Bases integration support.
- 1.4.2 Separate ingestion resources from document storageConfigure knowledge-base storage and data-source ingestion as distinct resources, then verify that synchronized documents are retrievable.
- 1.4.3 Align ingest and query embedding contractsConnect OpenSearch neural search to Amazon Bedrock embeddings and keep ingest, index, and query configurations compatible.
- 1.4.4 Keep metadata authority separate from vectorsCombine authoritative RDS or DynamoDB metadata, versioned S3 documents, and a rebuildable vector index without confusing their consistency boundaries.
- 1.4.5 Design provenance that survives document chunkingDesign typed document metadata that preserves authorship, time, domain, tenant, and access-policy identity through ingestion and retrieval.
- 1.4.6 Enforce tenant access before model contextBuild trusted retrieval filters and verify document permissions before authorized evidence enters a foundation model request.
- 1.4.7 Balance vector relevance against retrieval costChoose vector distance metrics and index designs using measured relevance, latency, memory, and operating cost.
- 1.4.8 Scale retrieval without multiplying search workDesign OpenSearch shard, replica, multi-index, and hierarchical retrieval layouts that balance throughput, fan-out, and recall.
- 1.4.9 Match source ingestion to sync ownershipChoose Knowledge Bases connectors, Kendra synchronization, or AppFlow transfers according to source type and the retrieval system you need to maintain.
- 1.4.10 Detect changes before scheduling index workUse DynamoDB Streams and scheduled repository refreshes to detect source changes without confusing detection with completed indexing.
- 1.4.11 Reconcile vectors against authoritative source revisionsMaintain retrieval consistency by reconciling incremental embeddings, deletions, stale revisions, and document-level ingestion failures.
- LABBuild Production-Ready Generative AI Applications on AWSConfigure Bedrock knowledge bases with S3 sources and OpenSearch Serverless storage for a multi-agent coffee shop application.
Design retrieval mechanisms for FM augmentation
- 1.5.1 Match chunk boundaries to answer evidenceChoose a Knowledge Bases chunking strategy that preserves useful evidence while balancing retrieval precision, context size, and ingestion cost.
- 1.5.2 Preserve evidence through parsing and transformationsConfigure Knowledge Bases parsers and custom Lambda transformations with explicit extraction, chunk, metadata, and S3 file contracts.
- 1.5.3 Choose embeddings with representative retrieval evidenceCompare Titan and other embedding models using language coverage, domain retrieval quality, vector dimensions, input behavior, and complete operating cost.
- 1.5.4 Keep batch embeddings compatible with queriesDesign bounded Lambda embedding batches that validate model identity, vector dimensions, representation, and index similarity before publication.
- 1.5.5 Configure stores for the retrieval contractConfigure OpenSearch and Aurora pgvector stores with compatible vector schemas, search indexes, permissions, and Knowledge Bases field mappings.
- 1.5.6 Place the boundary before evidence becomes contextChoose a Knowledge Bases retrieval API or custom pipeline according to who must inspect evidence, assemble context, and control generation.
- 1.5.7 Tune hybrid retrieval before accepting evidenceCombine lexical and semantic retrieval, verify backend support, and calibrate relevance thresholds without confusing them with authorization.
- 1.5.8 Require measured gains before enabling rerankingEvaluate Bedrock reranking against a fixed retrieval baseline and retain it only when relevance gains justify complete response latency and cost.
- 1.5.9 Transform queries without changing their intentChoose rewriting, expansion, or decomposition and bound Bedrock, Lambda, and Step Functions processing around the original retrieval intent.
- 1.5.10 Expose one retrieval contract across callersExpose authorized retrieval through a stable API, Bedrock function calling, and MCP clients while preserving validation and evidence provenance.
- 1.5.11 Pack authorized evidence with traceable citationsAssemble permission-checked retrieval results into a model-specific token budget while preserving source identity and verifying generated citations.
- LABBuild and Evaluate RAG Applications using Knowledge Bases for Amazon BedrockBuild Knowledge Bases question-answering applications with Retrieve and RetrieveAndGenerate, then evaluate them with RAGAS.
Implement prompt engineering strategies and governance for FM interactions
- 1.6.1 Separate instructions from enforceable response boundariesAssign model behavior, response structure, and safety enforcement to separate controls with explicit failure handling.
- 1.6.2 Freeze reusable prompts before production invocationBuild parameterized Bedrock prompts and distinguish editable drafts, saved variants, and version-qualified runtime references.
- 1.6.3 Promote reviewed prompts with a return pathRelease approved prompt configurations across environments and restore a known-good version without changing the reviewed artifact.
- 1.6.4 Join prompt changes to observed runtime usageCombine versioned S3 artifacts, CloudTrail activity, and CloudWatch invocation evidence to trace prompt ownership and use.
- 1.6.5 Persist conversations and clarify uncertain intentDesign DynamoDB conversation records and route intent decisions through Comprehend classification and bounded Step Functions clarification.
- 1.6.6 Test prompts against explicit acceptance contractsBuild prompt test suites that distinguish expected behavior, boundary handling, adversarial resistance, and regressions.
- 1.6.7 Refine prompt structure from measured failuresImprove structured inputs, output instructions, and reasoning patterns using controlled comparisons and independent evaluation.
- 1.6.8 Connect prompt stages with explicit data contractsBuild sequential Bedrock flows with typed connections, traceable intermediate results, and versioned runtime invocation.
- 1.6.9 Route flow branches through validated processingImplement ordered Bedrock flow conditions and reusable prompt and Lambda resources with explicit processing contracts.
- LABCreate Advanced AI Agents with Strands SDK and Amazon BedrockBuild an agent with the Strands SDK and Amazon Bedrock, and persist its conversation state.
Implementation and Integration
26% of the exam0 of 59 items completed
Implement agentic AI solutions and tool integrations
- 2.1.1 Bound the loop before granting toolsDesign a Strands agent loop with explicit model selection, tool contracts, working state, and enforceable termination conditions.
- 2.1.2 Choose who owns routing and delegationChoose Strands Agents and Agent Squad coordination patterns by request scope, context boundaries, and ownership of the final result.
- 2.1.3 Separate state lifetimes and isolation boundariesSeparate agent working state, Runtime sessions, and AgentCore Memory while binding all state access to trusted user and tenant identities.
- 2.1.4 Deploy agents with separate access boundariesDeploy agents on AgentCore Runtime and distinguish Gateway tool access from Identity credential management and end-user authorization.
- 2.1.5 Put mandatory steps under workflow controlChoose Step Functions, autonomous agent planning, or a hybrid architecture from ordering, recovery, approval, and failure requirements.
- 2.1.6 Bound reasoning cycles with explicit workflow stateDesign Step Functions workflows that bound model reasoning and tool actions with validated transitions, shared counters, and explicit terminal outcomes.
- 2.1.7 Enforce limits outside the agent promptCombine shared execution budgets, task deadlines, circuit-breaker state, and scoped IAM permissions to constrain agent actions.
- 2.1.8 Validate specialist answers before combining decisionsChoose model routing or specialist ensembles and resolve conflicting outputs through explicit evidence checks and aggregation policies.
- 2.1.9 Bind human approval to the exact actionDesign authenticated approval interfaces that resume Step Functions callbacks only after a reviewer authorizes an unchanged high-impact action.
- 2.1.10 Validate tool contracts at every boundaryDefine typed Strands tools and Lambda adapters that check parameters, validate outputs, and distinguish transport failures from business errors.
- 2.1.11 Preserve action identity across tool retriesMake side-effecting tools retry-safe with durable action identities, bounded recovery, and explicit compensating actions.
- 2.1.12 Match MCP clients to tool contractsIntegrate agents with MCP tools by matching protocol versions, discovering schemas, and validating invocation results.
- 2.1.13 Choose MCP hosting by state lifetimeChoose Lambda or ECS for MCP tools by matching request duration, state lifetime, resource needs, and recovery responsibilities.
- LABDeploying Intelligent Agents with Amazon Bedrock AgentCore RuntimeBuild Strands agents with custom tools and conversation management, then deploy a collaborating agent system with AgentCore.
- LABDeveloping Serverless Applications with AWS Step FunctionsBuild and run a Step Functions state machine that orchestrates a serverless workflow.
Implement model deployment strategies
- 2.2.1 Buy capacity for the measured demandChoose Bedrock on-demand inference, Provisioned Throughput, or SageMaker AI hosting from demand, capacity requirements, and operating cost.
- 2.2.2 Match inference modes to caller deadlinesSelect SageMaker AI inference modes by response timing, request size, hardware requirements, and the unit of offline work.
- 2.2.3 Join model paths through explicit contractsCombine Bedrock model invocations and separately hosted foundation models while preserving each path's interface, authorization, and deployment contract.
- 2.2.4 Bound Lambda callers before invoking modelsDesign Lambda model callers with separate initialization, deadline, payload, and concurrency budgets that respect downstream inference capacity.
- 2.2.5 Size containers beyond the model weightsSize SageMaker AI model containers using load readiness, GPU memory headroom, context-dependent cache demand, and measured generation throughput.
- 2.2.6 Assign inference operations before choosing hostingChoose SageMaker AI, ECS, EKS, or EC2 by matching inference requirements to the operational responsibilities your team can own.
- 2.2.7 Check compilation fit before tuning loadingDecide whether SageMaker Neo supports your model and hardware, then optimize artifact loading independently of compilation.
- 2.2.8 Prove smaller models meet the taskSelect and deploy task-specific JumpStart models when measured quality, endpoint cost, and latency improve the application outcome.
- 2.2.9 Escalate only when routine inference failsDesign bounded model cascades and distinguish application acceptance gates from Bedrock intelligent prompt routing.
- LABDeploy and Scale Containerized Applications with AWS FargateDeploy a containerized application on AWS Fargate behind a load balancer and scale it.
Design and implement enterprise integration architectures
- 2.3.1 Expose legacy operations through authenticated boundariesDesign authenticated enterprise adapters that preserve business authorization, protocol contracts, and legacy failure semantics.
- 2.3.2 Acknowledge webhooks after accepting durable workIntegrate FM-enabled microservices with authenticated API Gateway routes and Lambda handlers that distinguish acceptance from completion.
- 2.3.3 Choose integration contracts from failure semanticsChoose synchronous calls, asynchronous events, or synchronized data according to completion, recovery, and freshness requirements.
- 2.3.4 Select transports by delivery and replaySelect EventBridge, SNS, SQS, Kinesis, or Amazon MSK by routing, ordering scope, replay ownership, and throughput constraints.
- 2.3.5 Match source synchronization to change semanticsChoose CDC, DataSync, or Transfer Family and preserve revisions, deletions, and acceptance evidence before knowledge ingestion.
- 2.3.6 Separate workforce sign-in from service permissionsChoose enterprise federation and define separate least-privilege permissions for people, applications, and service workloads.
- 2.3.7 Authorize each user before agent accessEnforce verified end-user permissions on tools and retrieved evidence independently of the agent's AWS service role.
- 2.3.8 Verify residency across every processing routeDesign regional and hybrid GenAI routes by verifying compute, storage, logging, and recovery locations separately.
- 2.3.9 Release GenAI artifacts with tested rollbackBuild a gated release path that versions code, prompts, infrastructure, and deployment artifacts as a compatible package.
- 2.3.10 Centralize inference policy behind a gatewayDesign a GenAI gateway that authenticates callers, enforces routing policy, bounds usage, and correlates operational evidence.
- LABCreate and Manage a Basic Kafka Cluster using Amazon Managed Streaming for Apache Kafka (MSK)Create an MSK cluster and topic, send and view messages, and inspect basic cluster monitoring in the console.
- LABSecure AI Agent Access with Amazon Bedrock AgentCore IdentityDeploy an AgentCore agent with inbound authentication and test access using Cognito-issued JWTs.
Implement FM API integrations
- 2.4.1 Connect compute through explicit invocation contractsChoose a Bedrock SDK operation and configure credentials, model targeting, and response handling for the compute environment.
- 2.4.2 Track inference jobs across worker retriesDesign SQS inference workers with durable completion records, stable correlation IDs, and idempotent result publication.
- 2.4.3 Validate both sides of model inferenceCombine API Gateway request validation with application checks that reject incomplete, invalid, or unauthorized model results.
- 2.4.4 Preserve stream events through client deliveryConnect Bedrock streaming APIs to WebSocket, SSE, or chunked responses with explicit completion, error, and connection handling.
- 2.4.5 Bound SDK retries before switching modelsConfigure jittered SDK retries and coordinate attempt limits, deadlines, and approved fallback without duplicating visible results.
- 2.4.6 Protect inference with server-side request throttlesChoose API Gateway throttling and usage-plan controls that protect inference independently of client retry behavior.
- 2.4.7 Choose routing signals behind one contractSelect static, content-based, or metric-driven model routing while preserving a stable application request and response contract.
- 2.4.8 Put specialized routing in the right layerImplement specialized inference routes with application adapters, Step Functions branches, or bounded API Gateway transformations.
- 2.4.9 Join inference failures across service boundariesCorrelate model and tool failures using X-Ray trace context, API Gateway request identifiers, and per-attempt application records.
- LABDevelop Conversation Pattern with Amazon Bedrock APIsUse Nova Lite through the Bedrock API to compare zero-shot and contextual question answering with complete and streaming responses.
- LABIntegrate Amazon Bedrock Foundation Models into a Serverless AppBuild a serverless flash-card application with Amazon Bedrock, Lambda, and API Gateway.
- LABConfigure and Manage Message Queues with Amazon Simple Queue Service (SQS)Connect SQS queues to Lambda, configure SNS fan-out with message filtering, and handle failed messages with dead-letter queues.
Implement application integration patterns and development tools
- 2.5.1 Make generation outcomes explicit in APIsDesign a generation API that distinguishes partial output, token exhaustion, timeouts, and safe retry boundaries.
- 2.5.2 Connect accessible interfaces to explicit contractsCombine Amplify UI accessibility mechanisms with an OpenAPI-first contract that exposes generation progress and failure clearly.
- 2.5.3 Choose workflows by their operating ownerChoose Bedrock Flows, Amazon Quick, or existing Amazon Q Apps environments according to workflow ownership, audience, and service availability.
- 2.5.4 Extend conversations through bounded business adaptersAssign Lambda, Lex, Connect, and existing Q Business integrations distinct roles in conversational CRM workflows.
- 2.5.5 Separate document extraction from workflow acceptanceOrchestrate Bedrock Data Automation jobs with Step Functions while preserving invocation identity, output validation, and bounded recovery.
- 2.5.6 Validate assisted changes against measurable contractsUse Amazon Q Developer and Kiro to generate, refactor, test, and optimize GenAI application code with independently verified acceptance evidence.
- 2.5.7 Assign reasoning and execution separate boundariesCombine Strands Agents, Agent Squad, and Step Functions by assigning specialist routing, model reasoning, and durable execution distinct responsibilities.
- 2.5.8 Compose prompt chains through validated node contractsBuild reusable Amazon Bedrock prompt chains with explicit node inputs, validation gates, flow versions, and observable completion.
- 2.5.9 Diagnose integration failures from correlated runtime evidenceCombine CloudWatch Logs Insights, X-Ray call timing, and scoped Q Developer error analysis to isolate and verify GenAI integration failures.
- LABEnhance Production Applications with AI-assisted Development in KiroUse Kiro steering, powers, hooks, and a custom subagent to add frontend features and automate quality checks.
AI Safety, Security, and Governance
20% of the exam0 of 41 items completed
Implement input and output safety controls
- 3.1.1 Match guardrail policies to safety boundariesConfigure distinct Bedrock Guardrails policies for harmful content, prohibited topics, word matches, sensitive information, and source-grounded answers.
- 3.1.2 Place guardrail checks before content releaseChoose inline or independent Guardrails enforcement and handle intervention, masking, streaming exposure, and assessment failures explicitly.
- 3.1.3 Enforce moderation before processing and publicationCombine Comprehend toxicity signals, foundation-model assessments, Lambda adapters, and Step Functions decisions into bounded moderation paths.
- 3.1.4 Separate response shape from answer correctnessEnforce Bedrock JSON response contracts while retaining independent checks for business rules, evidence, authorization, and safe consumption.
- 3.1.5 Approve generated SQL before Athena executionValidate generated SQL structure, caller permissions, query scope, and resource controls before executing an unchanged approved query in Athena.
- 3.1.6 Verify answer claims against retrieved evidenceCombine Knowledge Bases passages, citation mappings, grounding confidence, and semantic similarity without mistaking relevance for factual support.
- 3.1.7 Separate tool defects from hostile instructionsClassify malformed tool responses and embedded instructions separately before admitting tool output into model context.
- 3.1.8 Detect attacks at every instruction boundaryDistinguish direct injection, indirect injection, and jailbreaks and place detection where untrusted content enters model context.
- 3.1.9 Assign each safety layer one gateCompose sanitization, classifiers, Guardrails, output validation, and response filtering with explicit failure and release decisions.
- 3.1.10 Measure bypasses alongside legitimate request failuresEvaluate safety controls on labeled adversarial and benign cases while separating detector errors, coverage gaps, and successful bypasses.
- 3.1.11 Bound hostile traffic before inference beginsProtect supported public GenAI front doors with AWS WAF rules, explicit inspection limits, and tested request-rate controls.
- LABSecure Web Applications with AWS WAFProtect an API Gateway API with an AWS WAF web ACL, IP and rate rules, and a managed SQL-injection rule set, then review the metrics.
Implement data security and privacy controls
- 3.2.1 Constrain private paths to approved APIsDesign private model and data API paths with endpoint policies that constrain callers, operations, and resources.
- 3.2.2 Authorize requests before trusting model instructionsSeparate AWS policy evaluation from application entitlement checks and enforce both before exposing data or executing tools.
- 3.2.3 Review grants with the right evidenceUse IAM Access Analyzer to distinguish external exposure, unused permissions, and policy defects before narrowing grants.
- 3.2.4 Constrain lake queries before assembling contextApply Lake Formation data filters through supported query paths and correlate permission changes with data-access evidence.
- 3.2.5 Separate data encryption from credential deliveryAssign KMS key permissions, encrypt application payloads with the AWS Encryption SDK, and retrieve credentials through Secrets Manager.
- 3.2.6 Detect personal data at each boundaryChoose Macie, Comprehend, and Bedrock Guardrails for source discovery and runtime PII detection without assuming complete coverage.
- 3.2.7 Preserve useful context while removing identifiersChoose privacy transformations by required disclosure, reversibility, and measured impact on foundation model utility.
- 3.2.8 Verify privacy before enabling invocation logsDistinguish Bedrock model privacy and retention controls from explicitly enabled invocation logging and customer-managed storage.
- 3.2.9 Sanitize logs before exporting agent tracesProtect application logs and agent traces with field-level sanitization and CloudWatch data protection independently of Bedrock invocation logging.
- 3.2.10 Delete retained data across every copyCombine S3 Lifecycle expiration with version-aware deletion and application workflows that verify removal of derived and logged data.
- LABConfiguring IAM Identity and resource-based policiesManage permissions with IAM identity and resource-based policies and set a role permissions boundary.
- LABSecure and Rotate Secrets with AWS Secrets ManagerCreate and retrieve a Secrets Manager secret, manage its authorization, and schedule Lambda-based rotation.
- LABSecure Sensitive Data with Amazon MacieConfigure Macie discovery jobs, protect findings with KMS, and analyze sensitive-data findings.
- LABMulti-part Uploads, Batch Operations, and Cross-Region Replication with S3Configure S3 Lifecycle deletion, inventory reports, batch jobs, and cross-Region replication.
Implement AI governance and compliance mechanisms
- 3.3.1 Turn governance requirements into tested controlsTranslate approved obligations into AWS enforcement mechanisms, test evidence, and risk entries with accountable owners.
- 3.3.2 Document the boundaries of model approvalUse SageMaker AI model cards to record intended use, limitations, evaluation evidence, and organizational approval without confusing documentation with deployment authorization.
- 3.3.3 Carry source lineage into answer evidenceConnect Glue Data Catalog metadata to versioned source and transformation records, then preserve evidence attributes in generated-response attribution.
- 3.3.4 Reconstruct decisions across audit evidence layersCombine CloudTrail API events, CloudWatch Logs decision records, and retained source attribution to reconstruct a GenAI response during forensic review.
- 3.3.5 Bind deployment approval to accountable ownersAssign governance owners and enforce deployment approval with versioned evidence, scoped permissions, and an auditable release decision.
- 3.3.6 Detect governance changes against explicit baselinesDefine separate baselines and measurements for misuse, answer quality, bias, and policy compliance before routing governance alerts.
- 3.3.7 Check eligibility before choosing drift checksChoose the appropriate SageMaker Model Monitor check while verifying customer eligibility, endpoint topology, data representation, and required labels.
- 3.3.8 Automate governance without losing decision evidenceConnect policy filtering, sensitive-text masking, protected response logs, alerts, and bounded remediation into an auditable governance workflow.
- LABImplement Logging and Monitoring using AWS CloudTrail and Amazon CloudWatchEnable CloudTrail logging, deliver it to CloudWatch, and build an alarm or metric filter from the resulting data.
Implement responsible AI principles
- 3.4.1 Explain answers through evidence and boundariesPresent verifiable answer explanations and source attribution while keeping private reasoning and unauthorized evidence out of public responses.
- 3.4.2 Explain agent actions through selected trace eventsUse Bedrock agent traces to reconstruct observable tool execution while distinguishing requests, results, private rationale, and business completion.
- 3.4.3 Measure uncertainty through calibrated evidence signalsPublish uncertainty signals to CloudWatch and interpret grounding scores, model self-confidence, and reviewed outcomes without conflating them.
- 3.4.4 Define fairness slices before choosing metricsDesign representative fairness datasets and select SageMaker Clarify bias metrics with explicit labels, facets, and service-access constraints.
- 3.4.5 Compare prompt fairness with calibrated judgesCompare versioned Bedrock prompts and flows across fairness slices, using LLM judges validated against human ratings.
- 3.4.6 Enforce approved use through explicit policy gatesCombine Bedrock Guardrails, versioned model-card evidence, and Lambda compliance decisions into enforceable responsible AI boundaries.
Operational Efficiency and Optimization for GenAI Applications
12% of the exam0 of 29 items completed
Implement cost optimization and resource efficiency strategies
- 4.1.1 Attribute inference spend to accepted workTrack model usage per invocation attempt and allocate inference cost to requests, features, and tenants.
- 4.1.2 Reduce tokens without discarding required meaningCombine context pruning, prompt compression, and output limits while measuring savings against accepted response quality.
- 4.1.3 Assign model tiers from measured evidenceRoute requests among eligible model tiers using capability gates, measured quality, latency, and cost per accepted result.
- 4.1.4 Choose inference delivery by deadline slackChoose batch or on-demand inference from request compatibility, deadline slack, measured queueing, and available capacity.
- 4.1.5 Match purchased capacity to usage trendsUse observed token demand, deadline slack, and utilization to choose Provisioned Throughput and batching policies without assuming automatic resizing.
- 4.1.6 Choose the right level of cache reuseDistinguish reusable Bedrock prompt prefixes from exact and semantic response caches by invocation behavior, identity, and correctness risk.
- 4.1.7 Bind cached answers to complete request identityDesign exact-response keys, request hashes, result fingerprints, and invalidation rules that preserve answer dependencies and freshness.
- 4.1.8 Verify semantic matches before reusing answersCalibrate semantic-cache thresholds and validate cached answers against the incoming question, authorized context, and current evidence.
- 4.1.9 Authorize cache reuse within explicit partitionsPartition response caches by trusted tenant, user entitlements, and answer context while enforcing permissions before every cached response is released.
- 4.1.10 Cache at the edge within privacy boundariesChoose CloudFront cache policies and Lambda@Edge event placement only when response variation, authorization, and privacy permit shared edge reuse.
- LABUsing Amazon ElastiCache to Implement CachingCreate and access an ElastiCache cluster, load Redis sample data, and read sorted-set scores.
Optimize application performance
- 4.2.1 Locate delay on the request pathAttribute response latency to model calls, retrieval, orchestration, and network boundaries before selecting an optimization.
- 4.2.2 Shorten the wait for useful outputCombine streaming, reusable precomputed work, and independent parallel branches to reduce response delay without bypassing acceptance gates.
- 4.2.3 Prove optimized inference earns its placeBenchmark Bedrock latency-optimized inference against standard inference with explicit quality, cost, model-support, and fallback evidence.
- 4.2.4 Tune retrieval speed behind relevance gatesImprove retrieval latency through measured index, query, and hybrid-scoring changes while preserving relevant authorized evidence.
- 4.2.5 Bound demand before increasing inference throughputImprove accepted inference throughput by coordinating worker concurrency, model quotas, batch eligibility, and token reservations.
- 4.2.6 Prove sampling changes improve accepted responsesTune supported model sampling controls and validate response quality, latency, and token use through controlled A/B comparisons.
- 4.2.7 Scale hosted capacity from measured demandPlan foundation model capacity and configure SageMaker scaling policies using backlog, concurrent work, token demand, and utilization evidence.
Implement monitoring systems for GenAI applications
- 4.3.1 Connect dashboards to individual decision pathsDesign correlated operational, model, retrieval, agent, and business telemetry that explains an individual request and its outcome.
- 4.3.2 Bound invocation logs by diagnostic purposeConfigure Bedrock invocation logging with supported destinations, delivery checks, and independently enforced retention.
- 4.3.3 Read runtime metrics within their boundariesTrack Bedrock token usage and latency with explicit endpoint, statistic, dimension, and invocation-counting boundaries.
- 4.3.4 Publish quality scores with explicit denominatorsTranslate evaluated prompt effectiveness, hallucination labels, and response quality into interpretable custom CloudWatch metrics.
- 4.3.5 Separate operational alerts from billing anomaliesDetect token bursts and measured quality drift with CloudWatch while using AWS Cost Anomaly Detection for delayed spend investigation.
- 4.3.6 Preserve request histories that explain changesBuild correlated request histories that connect user actions, configuration changes, and observed model behavior to protected investigation evidence.
- 4.3.7 Trace tools across agent handoffsReconstruct tool attempts and agent delegation paths to distinguish retries, execution failures, and anomalous usage.
- 4.3.8 Monitor retrieval health beyond query speedCombine vector-store metrics, ingestion evidence, freshness checks, and quality tests to evaluate retrieval health and optimization changes.
- 4.3.9 Diagnose failures with controlled output comparisonsUse reviewed golden cases, paired output comparisons, and sanitized execution evidence to isolate model and application failures safely.
- LABMonitor Agent Performance and Telemetry using AgentCore ObservabilityAnalyze AgentCore metrics and logs, diagnose failures, redeploy configuration changes, and configure alarms.
- LABMonitoring Amazon EC2 instance with Amazon Managed GrafanaConnect Amazon Managed Grafana to a CloudWatch data source and build a dashboard against an EC2 instance's metrics.
Testing, Validation, and Troubleshooting
11% of the exam0 of 19 items completed
Implement evaluation systems for GenAI
- 5.1.1 Define quality gates before selecting scoresTranslate factual accuracy, fluency, relevance, and output consistency into separate evaluation rubrics and acceptance gates.
- 5.1.2 Build datasets around production acceptance decisionsConstruct representative evaluation datasets with explicit acceptance criteria, risk slices, and reproducible Bedrock input records.
- 5.1.3 Compare models through accepted business outcomesCombine Bedrock evaluation results with measured token usage, latency, and business outcomes to choose a model under explicit constraints.
- 5.1.4 Resolve evaluator disagreement through reviewed evidenceInterpret automatic, judge-model, and human scores within their limits and route consequential disagreements to evidence-based review.
- 5.1.5 Separate causal experiments from deployment safetyDesign controlled model and prompt comparisons with stable cohort assignment, bounded canary exposure, and explicit promotion and rollback criteria.
- 5.1.6 Turn feedback into traceable reviewed labelsCollect response-linked user feedback and human annotations while preserving review context, workflow ownership, and service eligibility boundaries.
- 5.1.7 Gate releases on complete evaluation evidenceAutomate repeatable GenAI evaluations and regression decisions with explicit evidence, failure handling, and release gates.
- 5.1.8 Measure retrieval before judging generated answersEvaluate retrieval relevance, evidence coverage, context matching, and retrieval latency independently from answer quality.
- 5.1.9 Evaluate agent outcomes across execution levelsAssess agent completion, tool decisions, and multi-step behavior with AgentCore session, trace, and tool-level evaluation evidence.
- 5.1.10 Report comparisons with decision-ready evidenceBuild stakeholder dashboards that explain model experiment outcomes, acceptance gates, uncertainty, and operational trade-offs.
- 5.1.11 Validate deployed journeys and answer behaviorValidate GenAI deployments with synthetic journeys, controlled output-drift comparisons, and source-grounded hallucination regression tests.
- LABCreate Interactive Dashboards with Amazon Quick SightConnect a data source in Amazon QuickSight and build an interactive dashboard from it.
Troubleshoot GenAI applications
- 5.2.1 Locate context loss before changing modelsDiagnose context overflow and silent truncation by reconciling token counts, allocation decisions, and retained evidence.
- 5.2.2 Classify inference failures before retrying themDistinguish authorization, throttling, timeout, validation, and parsing failures so each receives a targeted diagnostic response.
- 5.2.3 Isolate output regressions with controlled casesDebug unexpected model behavior by controlling prompt cases, comparing frozen versions, and refining one tested cause at a time.
- 5.2.4 Locate answer defects across retrieval boundariesSeparate retrieval and generation failures through reviewed source relevance, compatible embedding contracts, and controlled drift comparisons.
- 5.2.5 Repair retrieval without hiding evidence defectsRemediate incompatible embeddings, stale source projections, broken chunks, and slow vector searches with independently verified retrieval repairs.
- 5.2.6 Trace template confusion through response contractsUse correlated logs and traces to isolate prompt-template defects and detect output-format regressions with schema and business-rule validators.
- 5.2.7 Prove fixes against representative regression casesBuild release evidence that a troubleshooting fix resolves the incident without degrading other supported behaviors or operational boundaries.