GenAIIDP Architecture

Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved. SPDX-License-Identifier: MIT-0

GenAIIDP Architecture

Flow Overview

1. Documents uploaded to Input S3 bucket trigger EventBridge events
2. Queue Sender Lambda records event in tracking table and sends to SQS
3. Queue Processor Lambda:
   • Picks up messages in batches
   • Manages workflow concurrency using DynamoDB counter
   • Starts Step Functions executions
4. Step Functions workflow runs the steps defined in the selected pattern to process the document and generate output in the Output S3 bucket
5. Workflow completion events update tracking and metrics

Components

• Storage: S3 buckets for input documents and JSON output
• Message Queue: Standard SQS queue for high throughput
• Functions: Lambda functions for queue operations
• Step Functions: Document processing workflow orchestration
• DynamoDB: Tracking and concurrency management
• CloudWatch: Comprehensive monitoring and logging
• Web UI: Browser-based interface for document management and visualization
  • CloudFront distribution for global availability (default), or ALB for VPC-based hosting (see ALB Hosting)
  • Cognito user authentication
  • GraphQL API for UI-backend interactions
• Evaluation: Document processing accuracy assessment system
• Document Knowledge Base: Optional Bedrock Knowledge Base for document querying

Modular Design Overview

The solution uses a modular architecture with nested CloudFormation stacks to support multiple document processing patterns while maintaining a common infrastructure for queueing, tracking, and monitoring. This design enables:

• Support for multiple processing patterns without duplicating core infrastructure
• Easy addition of new processing patterns without modifying existing code
• Centralized monitoring and management across all patterns
• Pattern-specific optimizations and configurations
• Optional features that can be enabled across all patterns:
  • Document summarization (controlled by configuration summarization.enabled property)
    • This feature also enables the “Chat with Document” functionality
    • This feature does not use the Bedrock Knowledge Base but stores a full-text text file in S3
  • Document Knowledge Base (using Amazon Bedrock)
  • Automated accuracy evaluation against baseline data

Stack Structure

Main Stack (template.yaml)

The main template handles all pattern-agnostic resources and infrastructure:

• S3 Buckets (Input, Output, Working, Configuration, Evaluation Baseline)
• SQS Queues and Dead Letter Queues
• DynamoDB Tables (Execution Tracking, Concurrency, Configuration)
• Lambda Functions for:
  • Queue Processing
  • Queue Sending
  • Workflow Tracking
  • Document Status Lookup
  • Evaluation
  • UI Integration and API Resolvers
• CloudWatch Alarms and Dashboard
• SNS Topics for Alerts
• Web UI Infrastructure:
  • CloudFront Distribution (default) or Application Load Balancer (ALB Hosting)
  • S3 Bucket for static web assets
  • CodeBuild project for UI deployment
• Authentication:
  • Cognito User Pool and Client
  • Identity Pool for secure AWS resource access
• AppSync GraphQL API for UI-backend communication

Pattern Stacks (patterns/*)

Each pattern is implemented as a nested stack that contains pattern-specific resources:

• Step Functions State Machine
• Pattern-specific Lambda Functions:
  • OCR Processing
  • Classification
  • Extraction
• Pattern-specific CloudWatch Dashboard
• Model Endpoints and Configurations

For detailed information about configuration capabilities, see configuration.md.

Unified Pattern Architecture

The solution uses a Unified Pattern that combines both BDA and pipeline processing modes into a single deployment. The use_bda configuration flag (set via the UI) controls which processing path is used at runtime:

• Pipeline mode (use_bda: false, default) — OCR → Classification → Extraction → Assessment → Rule Validation → Summarization → Evaluation
• BDA mode (use_bda: true) — BDA Invoke → BDA Process Results → Rule Validation → Summarization → Evaluation

Shared Processing Steps

Both modes share a common tail in the workflow:

• HITL Check — Routes documents to human review if confidence is below threshold
• Rule Validation — Applies configurable business rules (when enabled)
• Summarization — Generates document summaries using Bedrock LLM
• Evaluation — Compares results against ground truth baselines (when available)

Deployment

For detailed information on deploying this solution, see deployment.md.

The unified pattern is deployed as a single nested stack (PATTERNSTACK) containing all 12 Lambda functions for both processing modes. There is no pattern selector parameter — the processing mode is controlled entirely by the use_bda configuration flag set via the UI.

Note: The separate Pattern 1 and Pattern 2 deployments have been deprecated in favor of this unified architecture. See pattern-1.md and pattern-2.md for historical reference.

Integrated Monitoring

The solution creates an integrated CloudWatch dashboard that combines metrics from both the main stack and the selected pattern stack:

1. The main stack creates a dashboard with core metrics:

   • Queue performance
   • Overall workflow statistics
   • General error tracking
   • Resource utilization

2. Each pattern stack creates its own dashboard with pattern-specific metrics:

   • OCR performance
   • Classification accuracy
   • Extraction stats
   • Model-specific metrics

3. The DashboardMerger Lambda function combines these dashboards

For detailed information about monitoring capabilities, see monitoring.md.

Web UI Architecture

The solution includes a React-based web user interface for document management and visualization:

For detailed information about the Web UI, its features, and usage, see web-ui.md.

1. Authentication: Amazon Cognito provides secure user authentication and authorization
   • Admin users are created during deployment
   • Optional self-signup can be enabled with domain restrictions
   • Identity pools provide secure, temporary AWS credentials

• Content Delivery: CloudFront distribution serves the static web assets (default), or an Application Load Balancer for VPC-based hosting (see ALB Hosting)
  • Global availability and low latency (CloudFront) or private network access (ALB)
  • WAF integration for added security (optional)
  • Geographical restrictions can be applied

3. API Layer: AppSync GraphQL API connects the UI to backend services

   • Real-time data with subscriptions
   • Secure access control through Cognito and IAM
   • Lambda resolvers for complex operations

4. Document Operations: The UI supports:

   • Document upload and S3 presigned URL generation
   • Status tracking and results visualization
   • Configuration management
   • Knowledge base querying (when enabled)

Optional Features

Document Summarization

When enabled via the configuration summarization.enabled property (default: true), the solution provides document summarization across all patterns:

• All patterns use a dedicated summarization step with Bedrock models
• Summarization provides a concise overview of the document content
• Results can be viewed in the Web UI and downloaded or printed
• Configuration settings control summarization behavior per pattern

Document Knowledge Base

The solution optionally integrates with Amazon Bedrock Knowledge Base:

• Processed documents are indexed in a knowledge base
• Enables natural language querying of document content
• Supports various Bedrock models (Amazon Nova, Anthropic Claude)
• GraphQL API integration allows querying from the UI

For detailed information about the Knowledge Base integration, see knowledge-base.md.

Accuracy Evaluation

The solution includes a comprehensive evaluation system:

1. Baseline Data: Ground truth data stored in the evaluation baseline bucket
2. Automatic Evaluation: When enabled, each processed document is automatically evaluated against baseline data if available
3. Metrics:
   • Extraction accuracy for key-value pairs
   • Classification accuracy across document types
   • Summarization quality assessment
4. UI Integration: Results visualized in the web interface
5. CloudWatch Metrics: Aggregated accuracy metrics for monitoring

For detailed information about the evaluation system, see evaluation.md.

Bedrock Guardrails Integration

The solution supports optional Amazon Bedrock Guardrails integration:

• Define content boundaries for Bedrock model outputs
• Apply guardrails to all Bedrock model interactions across all patterns
• Support for both DRAFT and versioned guardrails
• Configuration parameters:
  • BedrockGuardrailId - ID of an existing Bedrock Guardrail
  • BedrockGuardrailVersion - Version of the guardrail to use (e.g., DRAFT, 1, 2)
• Guardrails can be applied to:
  • Extraction operations
  • Summarization (when enabled)
  • Knowledge Base interactions (when enabled)

Post-Processing Lambda Hook

The solution supports an optional post-processing Lambda hook integration:

• Automatically triggered via EventBridge after a document is successfully processed
• Configured via the PostProcessingLambdaHookFunctionArn parameter
• Enables custom downstream processing of document extraction results
• Can be used for integration with other systems such as:
  • Enterprise document management systems
  • Business process workflows
  • Data analytics pipelines
  • Custom notification systems
• Receives the document processing details and output location

For comprehensive implementation guidance, use cases, and code examples, see post-processing-lambda-hook.md.

Additional Documentation

• classification.md - Details on document classification capabilities
• extraction.md - Details on data extraction capabilities
• troubleshooting.md - Troubleshooting guidance and common issues