document_processing

Document Processing

Overview

The Document Processing use case automates the classification, data extraction, and compliance validation of financial documents. It handles loan applications, KYC documents, financial statements, regulatory filings, and contracts -- classifying each by type and jurisdiction, extracting structured fields, and validating against regulatory rules to surface compliance issues.

Business Value

• Accelerated intake -- documents classified and fields extracted in seconds, reducing manual processing backlogs
• Compliance assurance -- automated validation against regulatory rules catches missing fields and inconsistencies before human review
• Structured output -- unstructured documents transformed into standardized key-value pairs for downstream systems
• Jurisdiction awareness -- classification includes regulatory framework mapping for multi-jurisdictional operations
• Reduced error rates -- parallel agent cross-checks improve extraction accuracy over single-pass approaches

Architecture

graph TB
    Request["Client Request"] --> Runtime["AgentCore Runtime"]
    Runtime --> Orchestrator["Orchestrator"]
    Orchestrator --> Classifier["Document Classifier<br/><small>Type, jurisdiction & regulatory classification</small>"]
    Orchestrator --> Extractor["Data Extractor<br/><small>Structured data extraction</small>"]
    Orchestrator --> Validator["Validation Agent<br/><small>Compliance rule validation</small>"]
    Classifier --> Bedrock["Amazon Bedrock<br/>(Claude)"]
    Extractor --> Bedrock
    Validator --> Bedrock
    Classifier --> S3["S3 Sample Data"]
    Extractor --> S3
    Validator --> S3
    Classifier --> Synthesis["Result Synthesis"]
    Extractor --> Synthesis
    Validator --> Synthesis
    Synthesis --> Response["Response"]

Loading

Directory Structure

use_cases/document_processing/
├── README.md
└── src/
    └── strands/
        ├── __init__.py
        ├── config.py          # DocumentProcessingSettings
        ├── models.py          # Pydantic request/response models
        ├── orchestrator.py    # DocumentProcessingOrchestrator + run_document_processing()
        └── agents/
            ├── __init__.py
            ├── document_classifier.py
            ├── data_extractor.py
            └── validation_agent.py

Agentic Design

The orchestrator uses a parallel fan-out pattern. In full mode, all three agents run concurrently via asyncio.gather. Targeted modes (classification_only, extraction_only, validation_only) invoke a single agent. The orchestrator synthesizes combined results into a structured JSON summary with classification, extraction completeness, and validation status.

Agents

Agent	Role	Data Used	Output
Document Classifier	Classifies documents by type (loan application, KYC, financial statement, regulatory filing, contract), identifies jurisdiction and regulatory relevance	Document profile via s3_retriever_tool	Document type with confidence score, jurisdiction, applicable regulations
Data Extractor	Extracts structured data from unstructured financial documents including entities, amounts, dates, and reference numbers	Document profile via s3_retriever_tool	Key-value fields, named entities, financial amounts, dates, completeness assessment
Validation Agent	Validates extracted data against regulatory rules and business logic, cross-references with existing records	Document profile via s3_retriever_tool	Validation status (VALID/INVALID/REVIEW_REQUIRED), passed/failed checks, notes

Data and Tools

• Tool: s3_retriever_tool -- retrieves document profiles and content from S3
• S3 data prefix: samples/document_processing/
• Model: Claude Sonnet (via Amazon Bedrock), temperature 0.1, max 8192 tokens
• Config thresholds: classification_confidence_threshold=0.85, extraction_accuracy_threshold=0.90, max_document_size_mb=50

Request / Response

Request -- ProcessingRequest:

Field	Type	Description
document_id	str	Document identifier (e.g., DOC001)
processing_type	ProcessingType	full, classification_only, extraction_only, validation_only
additional_context	str | None	Optional context

Response -- ProcessingResponse:

Field	Type	Description
document_id	str	Document identifier
processing_id	str	Unique processing UUID
timestamp	datetime	Processing timestamp
classification	DocumentClassification | None	Document type, confidence, jurisdiction, regulatory relevance
extracted_data	ExtractedData | None	Fields dict, entities, amounts, dates
validation_result	ValidationResult | None	Status, checks passed/failed, notes
summary	str	Processing summary
raw_analysis	dict	Raw agent output

Quick Start

# Deploy to AgentCore
USE_CASE_ID=document_processing ./scripts/deploy/full/deploy_agentcore.sh

# Test the deployment
./scripts/use_cases/document_processing/test/test_agentcore.sh

Sample Data

Located at data/samples/document_processing/

Document ID	Type	Description
DOC001	Loan Application	Commercial loan application for Acme Corporation Ltd -- 12-page PDF, $5M business expansion loan, 5-year term, US jurisdiction

Related Documentation

• FSI Foundry Overview
• Architecture Patterns
• Deployment Guide
• Implementation Details