Amazon Bedrock/Overview

RCW guided build

Secure generative AI support assistant

Design, build, test, and approve a grounded assistant with governance controls from the first decision through operations.

Intermediate · 60–90 min

This is an interactive console simulation.

No AWS account, credentials, network calls, or real cloud charges are used. Product names appear only to teach the workflow; no vendor logo is used.

Scenario

Build the RCW IT policy assistant

Employees need reliable answers from approved IT policies. Your assistant must cite its source, resist prompt injection, protect personal data, and require a person to approve ticket creation.

✓ Select a cost-aware foundation model
✓ Apply Guardrails to input and output
✓ Create a Knowledge Base with synthetic policy documents
✓ Prepare an Agent with a gated action group
✓ Evaluate grounding, safety, authorization, and cost controls

Reference architecture

Governed retrieval and action flow

Simulated

Objective 1 · Govern and map

Record the use case and data boundary

Complete this intake before opening model access. This creates the scope and accountability evidence for the lab.

10 points

Learner name Business owner Highest permitted data class

Approved purposeIT policy question answeringGrounded, cited answers from approved internal policy content.Fully autonomous employment decisionsHigh-impact decisions without human review.

Required attestationsI will use only synthetic, non-sensitive training content.A human owner remains accountable for outputs and actions.I will re-assess impact, residency, and legal obligations before production.

Framework intent: NIST AI RMF Govern/Map and ISO/IEC 42001 Plan.

Amazon Bedrock/Model catalog

Foundation models

Model catalog

Compare available models against quality, latency, cost, licensing, and data-residency needs.

Not selected

Model governance checkpoint

Production teams should review model cards, acceptable-use requirements, provider terms, Region availability, EULAs, and cross-Region routing before enabling access.

Recommended for this lab

Amazon Nova Lite

Amazon

TextConverseOn-demand

Use case: Fast, cost-aware support Q&A
Relative latency: Low
Relative cost: $

Inference option

Amazon Nova Pro

Amazon

TextMultimodalOn-demand

Use case: Complex reasoning and analysis
Relative latency: Medium
Relative cost: $$$

Amazon Titan Text Premier

Amazon

TextLegacy comparison

Use case: General text generation
Relative latency: Medium
Relative cost: $$

Amazon Bedrock/Guardrails/Create

Build

Create guardrail

Apply consistent safety, privacy, and policy controls to prompts and model responses.

No draft

1 Configure2 Review3 Create and publish

Objective 2

Guardrail details

15 points

NameBlocked response

Content filters

Set filter strengths for both user inputs and generated responses.

Category	Input strength	Output strength
Hate
Insults
Sexual content
Violence
Misconduct
Prompt attacks		Input only

Denied topics

Block requests for secrets and attempts to bypass policy.

Credentials, passwords, tokens, and private keysRequests to bypass safeguards or reveal hidden instructions

Sensitive information filters

Choose how PII in model output is handled.

Email address — maskPhone number — blockAccess keys and passwords — block

Contextual grounding check

Require grounded responses when source context is suppliedGrounding threshold

Tip: guardrails are defense-in-depth, not a replacement for authorization or output validation.

Amazon Bedrock/Knowledge bases/Create

Build

Create knowledge base with vector store

Ingest approved documents, transform them into embeddings, and retrieve source-grounded context.

Not created

1 Details2 Data source3 Data storage4 Review

Objective 3

Knowledge Base details

20 points

NameIAM service roleData deletion policy

Data source

Source typeS3 URIData classification

PDF incident-response-policy.pdf 8 KB · approvedPDF access-control-standard.pdf 6 KB · approvedTXT remote-work-guidance.txt 4 KB · approved

Chunking, embeddings, and vector storage

Chunking strategyEmbeddings modelVector store

Customer managed KMS keyS3 Block Public AccessPrivate network path reviewed

Synthetic documents only. Real data requires access-control filtering, retention, residency, and legal review.

Amazon Bedrock/Agents/Agent builder

Build

Agent builder

Orchestrate the selected model, trusted knowledge, and a tightly scoped action with human oversight.

DRAFT

Objective 4

Agent details

15 points

Agent nameFoundation modelService role

Instructions for the Agent

Knowledge bases

rcw-it-policy-kbUse approved IT policy content for cited answers.rcw-it-policy-guardrail · version 1Apply to inputs and outputs.

Session security

Idle session timeoutUse customer managed KMS encryptionDo not retain conversation memory after session

Action group: Create support ticket

A simulated Lambda function can create a ticket. Keep the tool read/write scope narrow and require approval before execution.

Action schemaLambda execution roleConfirmation

Agent tools are security boundaries. Never rely on the model alone to authorize an action.

Amazon Bedrock/Playgrounds/Agent test

Test

Agent playground

Verify grounded answers, injection resistance, PII protection, and approval-gated actions.

0 / 4 tests

Test rcw-it-helpdesk-agentAlias: lab-v1 · synthetic session

Lab assistant

Prepare the model, guardrail, Knowledge Base, and Agent, then run all four required tests above.

Objective 6

Required test evidence

10 points

○

Grounded answerAnswer includes a source citation.

○

Prompt injectionRetrieved or direct hostile instructions are blocked.

○

PII disclosureSensitive data request is denied or masked.

○

Excessive agencyTicket executes only after human approval.

Amazon Bedrock/Settings/Logging and monitoring

Operate

Logging, monitoring, and cost controls

Create traceable evidence without retaining sensitive content longer than required.

Not configured

Invocation logs can contain full prompts and responses.

Payload logging is enabled here only because every prompt and document is synthetic. Production use requires privacy review, least-privilege access, encryption, destination controls, and documented retention.

Objective 5

Audit trail

10 points

CloudTrail multi-Region trailRecord management activity and deliver to protected S3 storage.Bedrock data-event selectorsAgentAlias, KnowledgeBase, Guardrail, and Flow resource types.Log file validation and protected archiveUse KMS encryption and object retention where required.

Model invocation logging

Text invocation logging — synthetic lab onlyPublish supported Converse/InvokeModel request and response records.DestinationCloudWatch Logs retentionEncrypt destinations with customer managed KMS key

Metrics, alarms, and budget

Invocation latency/error alarmGuardrail intervention alarmToken usage and budget alert

I will not put secrets, credentials, or personal data in tags, resource names, or request metadata.

CloudTrail data events and log destinations can add cost. Select only required events and set a retention policy.

Amazon Bedrock/Model evaluation/Create job

Measure

Create RAG and safety evaluation

Use a versioned benchmark to measure usefulness and safety before release.

No job

Amazon Bedrock/Compliance review

Manage and approve

Production-readiness review

Inspect the evidence, document residual risk, and make a human release decision.

Not ready

Control evidence

Automated readiness checks

0 / 8 ready

○

Governance and data scopeOwner, purpose, classification, and impact attestations recorded.

Pending

○

Model decisionCost-aware model and single-Region inference selected.

Pending

○

Guardrail versionContent, prompt attack, denied topic, PII, and grounding controls published.

Pending

○

Grounded knowledgeApproved data synced with encryption and deletion controls.

Pending

○

Least-agency AgentScoped tool, least privilege, session controls, and approval gate.

Pending

○

Audit and operationsEncrypted, retained logs plus security and budget alarms.

Pending

○

Adversarial testsGrounding, injection, PII, and excessive-agency tests passed.

Pending

○

Evaluation thresholdsVersioned RAG and safety evaluation passed with human review planned.

Pending

Secure generative AI support assistant

Build the RCW IT policy assistant

Governed retrieval and action flow

Record the use case and data boundary

Model catalog

Amazon Nova Lite

Amazon Nova Pro

Amazon Titan Text Premier

Create guardrail

Guardrail details

Content filters

Denied topics

Sensitive information filters

Contextual grounding check

Create knowledge base with vector store

Knowledge Base details

Data source

Chunking, embeddings, and vector storage

rcw-approved-policy-source

Agent builder

Agent details

Knowledge bases

Session security

Action group: Create support ticket

Agent playground

Required test evidence

Logging, monitoring, and cost controls

Audit trail

Model invocation logging

Metrics, alarms, and budget

Recent simulated events

Create RAG and safety evaluation

Evaluation details

Release thresholds

rcw-it-policy-eval-v1 results

Production-readiness review

Automated readiness checks

Human release decision