Review & Assessment Powered by Intelligent Documentation (RAPID)

Document	Language
README (this page)	English | 日本語
Deployment Options (CloudShell options, closed network, AI models)	English | 日本語
Developer Guide (architecture, troubleshooting)	English | 日本語
Local Development (run the app on your machine)	English | 日本語
Example Use Cases (industry sample scenarios and documents)	English | 日本語

This sample is a document review solution powered by generative AI (Amazon Bedrock). It streamlines review processes involving extensive documents and complex checklists using a Human in the Loop approach. It supports the entire process from checklist structuring to AI-assisted review and final human judgment, reducing review time and improving quality.

Important

This tool is intended only for decision support and does not provide professional judgment or legal advice. All final judgments must be made by qualified human experts.

Warning

This sample may undergo breaking changes without prior notice.

How It Works

RAPID performs document review in two phases:

1. Build a checklist – Upload a document (PDF) — such as a regulation, guideline, or specification — that describes what to check and where, and AI extracts the review criteria as a checklist.
2. Run a review – Upload the documents to be reviewed (PDF or images) and pick the checklist to compare them against, and AI evaluates each item as Pass / Fail, presenting a confidence score, the AI's rationale, and the documents it referenced.

RAPID runs on AWS serverless services (Amazon CloudFront, API Gateway + Lambda, Step Functions, Aurora Serverless v2, and Amazon Bedrock / AgentCore). See the Developer Guide for the architecture diagram.

Key Features

• AI checklist extraction – Converts regulations, guidelines, specifications, and the like into a checklist.
• AI document review – Judges each checklist item as Pass / Fail and presents a confidence score, the AI's rationale, and the documents it referenced.
• Per-checklist-item model selection – Assigns any generative AI model to each checklist item, so you can spend higher-cost models only on the difficult checks.
• Agent tools – Equips a checklist item with Amazon Bedrock Knowledge Bases (RAG), the AgentCore Code Interpreter (code execution for calculations and validation), and MCP (Model Context Protocol) servers when the check needs knowledge from external tools.
• Customizable prompts – Lets you review and edit the system prompts used for checklist extraction on a dedicated Prompt Management screen.
• Example use cases gallery – Ships with industry sample scenarios (real estate, IT, manufacturing, healthcare, corporate governance, and more), so you can try RAPID's document review right away.
• Closed / private network deployment – Runs RAPID without exposing it to the internet. Combined with AWS Site-to-Site VPN or AWS Direct Connect, you can use RAPID from your on-premises network over fully private connectivity. See Closed / Private Network Deployment.
• Concurrency control – Keeps reviews within Amazon Bedrock's quotas by controlling how many run concurrently.

Screenshots of the main screens (click to expand)

Key Use Cases

• Product specification compliance review – Verify that product specifications meet requirements and industry standards, and let reviewers concentrate on the final confirmation.
• Technical manual quality verification – Check that technical manuals comply with internal guidelines and industry standards, and detect missing information and inconsistencies automatically.
• Procurement document compliance verification – Extract the required information from procurement documents and proposals spanning hundreds of pages, and have humans verify the compliance results.

Concrete scenarios with sample documents are available in the examples gallery.

Deployment Methods

1. Deployment Using CloudShell (For Those Who Want to Start Easily)

This method allows you to deploy directly from your browser using AWS CloudShell, without preparing a local environment.

1. Enable Amazon Bedrock models

   Access Bedrock Model Access from the AWS Management Console and enable access to the models you plan to use (see AI Model Customization for the model list). By default, the Oregon (us-west-2) region is used for Amazon Bedrock, but you can change it with the --bedrock-region option.

2. Open AWS CloudShell

   Open AWS CloudShell in the region where you want to deploy.

3. Run the deployment script

   wget -O - https://raw.githubusercontent.com/aws-samples/review-and-assessment-powered-by-intelligent-documentation/main/bin.sh | bash

   This command automatically executes everything from repository cloning to deployment. Upon completion, the frontend URL and the API URL are displayed; open the frontend URL in your browser to start using the application.

4. Specify custom options (optional)

   wget -O - https://raw.githubusercontent.com/aws-samples/review-and-assessment-powered-by-intelligent-documentation/main/bin.sh | bash -s -- --ipv4-ranges '["192.168.0.0/16"]'

   Options such as --ipv4-ranges and --closed-network correspond to the CDK parameters described in Parameter Customization. For the full list of options, see CloudShell Deployment Options.

Important

With this deployment method, if you do not set option parameters, anyone who knows the URL can sign up. For production use, we strongly recommend adding IP address restrictions and disabling self-signup (--cognito-self-signup false).

2. Deployment from Local Environment (Recommended for Customization)

Note

This method requires Docker to be installed and running locally, because the CDK build bundles several Lambda functions as container images (Prisma database migration, review processor, AgentCore runtime). Node.js and AWS credentials with permissions for the target account / region are also required.

• Clone this repository:

git clone https://github.com/aws-samples/review-and-assessment-powered-by-intelligent-documentation.git
cd review-and-assessment-powered-by-intelligent-documentation

• Edit parameter.ts as needed. See Parameter Customization for details.
• Bootstrap the target region once before the first deployment. The exported AWS_DEFAULT_REGION applies to both the bootstrap and the deployment; you can also specify the region per command with npx cdk bootstrap aws://<account-id>/<region> instead. Run npm ci in cdk/ before bootstrapping: cdk bootstrap loads the CDK app in cdk/bin/rapid.ts, so on a fresh clone it fails before reaching AWS. The app also bootstraps us-east-1 for the CloudFront WAF stack (skipped in the S3 + API Gateway and closed-network frontend modes).

cd cdk
npm ci
export AWS_DEFAULT_REGION="<region>"
npx cdk bootstrap

• Deploy (this builds all packages and deploys them automatically):

cd cdk
npm run deploy

Manual step-by-step deployment

Prepare the backend:

cd backend
npm ci
npm run prisma:generate
npm run build

Then install the CDK packages and deploy:

cd ../cdk
npm ci
npx cdk deploy --require-approval never --all

• You will see output like the following. Access the Web application URL displayed in RapidStack.FrontendURL from your browser.

 ✅  RapidStack

✨  deployment time: 78.57s

Output:
...
RapidStack.FrontendURL = https://xxxxx.cloudfront.net

Cleaning Up (Destroying the Stacks)

To remove everything this sample created and stop incurring costs, destroy both CDK stacks. Run this from the cdk directory:

cd cdk
npx cdk destroy --all

--all lets CDK delete the stacks in dependency order (it removes RapidStack before RapidFrontendWafStack, which lives in us-east-1).

Warning

This is a sample/demo configuration: the S3 buckets, the Aurora database, and — when the stack created it — the Cognito User Pool are all removed on destroy, including all data and user accounts. There is no retention or deletion protection. Back up anything you need first, and for real workloads consider changing these removal policies.

A few resources are not removed automatically, such as an imported Cognito User Pool, some CloudWatch Logs log groups, and the RapidCodeBuildDeploy stack created by the CloudShell deployment. In VPC mode, cdk destroy can also fail temporarily with DELETE_FAILED while service-managed ENIs are released. See Cleanup Details for both topics.

Parameter Customization

The following parameters can be customized during CDK deployment. Edit cdk/lib/parameter.ts:

Parameter Group	Parameter Name	Description	Default Value
WAF Configuration	allowedIpV4AddressRanges	IPv4 ranges to allow in the frontend WAF	["0.0.0.0/1", "128.0.0.0/1"] (all allowed)
	allowedIpV6AddressRanges	IPv6 ranges to allow in the frontend WAF	["0000::/1", "8000::/1"] (all allowed)
Cognito Settings	cognitoUserPoolId	Existing Cognito User Pool ID	Create new
	cognitoUserPoolClientId	Existing Cognito User Pool Client ID	Create new
	cognitoDomainPrefix	Prefix for the Cognito domain	Auto-generated
	cognitoSelfSignUpEnabled	Whether to enable self-signup for the Cognito User Pool	true (enabled)
Migration	autoMigrate	Whether to automatically run database migration during deployment	true (auto-run)
MCP Features	mcpAdmin	Whether to grant admin permissions to the MCP runtime Lambda function	false (disabled)
Citations API	enableCitations	Whether to enable the Citations API for PDF documents (AWS announcement)	true (enabled)
Model Selection	availableModels	List of models available for per-checklist-item model selection. Set to an empty array [] to disable the model selection UI	Claude Opus 4.6, Claude Sonnet 4.6, Claude Haiku 4.5, Claude Sonnet 4
Network Mode	s3ApiGatewayFrontend	Serve the SPA from S3 via a dedicated REGIONAL API Gateway (S3 proxy) instead of CloudFront, keeping standard networking. See Closed / Private Network Deployment.	false
	closedNetwork	Fully private mode: isolated subnets, no NAT, VPC endpoints, PRIVATE API Gateways, Cognito PrivateLink. Implies s3ApiGatewayFrontend. See Closed / Private Network Deployment.	false
	agentCoreNetworkMode	AgentCore Runtime network mode (only applies when closedNetwork). PUBLIC = runtime has internet (MCP/uv work); VPC = runtime fully isolated. Invoke path is private either way	PUBLIC
Map State Concurrency	reviewMapConcurrency	Map State concurrency for the Review Processor (must be configured in consultation with throttling limits)	1
	checklistInlineMapConcurrency	Inline Map State concurrency for the Checklist Processor (must be configured in consultation with throttling limits)	1
Review Queue Settings	reviewMaxConcurrency	Max concurrent Step Functions executions for the review queue consumer	2
	reviewQueueMaxDepth	Max queue depth before the API returns a global concurrency limit error	10
	reviewQueueMaxQueueCountMs	Max wait time in ms before error handling in the review queue consumer	86,400,000 (24h)
	reviewQueueLogLevel	Log level for the review queue consumer Lambda	WARNING
Schedule Settings	feedbackAggregatorScheduleExpression	Feedback Aggregator execution schedule (EventBridge Scheduler expression format)	cron(0 2 * * ? *) (Daily at 2:00 UTC)

Schedule Expression Format:

• Cron format: cron(minute hour day month day-of-week year) - Example: cron(0 2 * * ? *) (Daily at 2:00 UTC)
• Rate format: rate(value unit) - Example: rate(1 day) (Every day), rate(12 hours) (Every 12 hours)
• Details: Schedule types on EventBridge Scheduler

Caution

The default values prioritize an easy trial over production hardening:

• WAF IP restrictions: the defaults allow all IP addresses. For production, set the specific IP ranges you want to allow.
• Self-signup is enabled by default. For production use, we strongly recommend setting cognitoSelfSignUpEnabled: false; leaving it enabled allows anyone who reaches the URL to register an account.
• autoMigrate runs database migrations automatically on every deployment. For production environments or environments containing important data, consider setting it to false and controlling migrations manually.

Closed / Private Network Deployment

Setting closedNetwork: true deploys RAPID in a fully closed configuration: the VPC has only isolated subnets (no NAT / Internet Gateway), runtime AWS access goes through VPC endpoints, and both API Gateways become PRIVATE endpoints. The application is then reachable only from inside the VPC — for example from your on-premises network connected via AWS Client VPN, AWS Site-to-Site VPN, or AWS Direct Connect. If you only need to avoid CloudFront while staying publicly reachable, use s3ApiGatewayFrontend: true instead.

Closed network mode comes with several constraints — deployment itself still requires internet access, authentication is limited to Cognito SRP, and toggling the mode on an existing stack replaces the VPC, among others. Be sure to read Closed / Private Network Deployment before enabling it.

AI Model Customization

RAPID uses Strands agents with tools such as file reading, so you must select models that support tool use. You can change the processing models (documentProcessingModelId / imageReviewModelId) and the per-item selection list (availableModels) in parameter.ts. For the list of tool-use capable models, notes on cross-region inference profiles, and configuration examples, see AI Model Customization.

Pricing

This solution incurs infrastructure fixed costs (~$5/day, ~$150/month, mainly for the NAT Gateway and Aurora Serverless v2) plus Amazon Bedrock usage costs based on document processing volume (pay-per-use).

Model class	Processable pages per review	Cost example
Budget-friendly lightweight model (Claude Haiku 4.5, etc.)	~80–85 pages	~$0.28 for 80 pages
High-accuracy large-capacity model (Claude Opus 4.6, etc.)	~430 pages	~$5.75 for 400 pages

Important

• Please test with your own sample documents to determine actual costs. Costs vary significantly with text volume, image count / size, and the number of checklist items (page counts are rough estimates only).
• Agent features (Knowledge Bases, Code Interpreter, etc.) may incur up to 10x higher costs.
• Detailed pricing and token usage can be viewed on the review results screen.
• The Amazon Bedrock Converse API has a 4.5 MB file size limit.

For the latest pricing information, please visit the Amazon Bedrock pricing page.

User Roles and Admin Setup

Role Behavior (Admin / General User)

• Admin: Can view and operate on all checklist sets and review jobs (no owner restriction).
• General user: Can access only resources they own (owner-restricted).

Resource	Owner	Action	Admin	General User
Checklist	Self-created	View	O	O
Checklist	Self-created	Edit	O	O
Checklist	Self-created	Delete	O	O
Checklist	Created by others	View	O	X
Checklist	Created by others	Edit	O	X
Checklist	Created by others	Delete	O	X
Review	Self-created	View	O	O
Review	Self-created	Edit	O	O
Review	Self-created	Delete	O	O
Review	Created by others	View	O	X
Review	Created by others	Edit	O	X
Review	Created by others	Delete	O	X

Admin Initial Setup

This project uses a Cognito custom attribute rapid_role. When the ID token contains custom:rapid_role=admin, the backend treats the user as an admin.

1. In the Cognito User Pool, set the custom attribute rapid_role to admin for the target user.
2. Confirm the ID token includes custom:rapid_role=admin after login.

For local development, setting RAPID_LOCAL_DEV=true makes requests run as an admin user.

Contact

• Takehiro Suzuki
• Kenta Sato

Contribution

See CONTRIBUTING for more information.

License

This project is distributed under the license described in LICENSE.