Skip to main content

​
Overview

The Document Extraction feature uses advanced Large Language Models (LLMs) to extract structured CDM-compliant data from unstructured PDF credit agreements. It supports multiple input modes including PDF upload, text input, audio transcription, and image OCR. Code Reference: client/src/apps/docu-digitizer/DocumentParser.tsx, app/chains/extraction_chain.py, app/api/routes.py (extract endpoints)

​
Key Features

​
Multi-Modal Input

  • PDF Upload: Direct PDF file upload and processing
  • Text Input: Paste or type document text
  • Audio Transcription: Speech-to-text transcription using Gradio Spaces
  • Image OCR: Optical character recognition from images
  • Document Library: Select from previously extracted documents
Code Reference: client/src/apps/docu-digitizer/MultimodalInputTabs.tsx

​
LLM Extraction

  • Simple Extraction: For documents under 50k characters
  • Map-Reduce Extraction: For larger documents (>50k characters)
  • Multiple Providers: OpenAI GPT-4o, vLLM, HuggingFace
  • Structured Output: CDM-compliant structured data extraction
Code Reference: app/chains/extraction_chain.py, app/chains/map_reduce_chain.py

​
CDM Compliance

  • Automatic Validation: Policy compliance validation at extraction point
  • CDM Models: CreditAgreement, Party, LoanFacility models
  • Event Generation: Automatic CDM event generation
  • Policy Enforcement: Real-time policy evaluation
Code Reference: app/models/cdm.py, app/models/cdm_events.py

​
Workflow

​
1. Input Document

  1. Choose Input Method: PDF upload, text paste, audio, or image
  2. Upload/Enter Content: Provide document content
  3. Review Input: Verify content is correct
  4. Proceed to Extraction: Click “Extract” button

​
2. Extraction Process

  1. LLM Processing: System processes document with selected LLM
  2. Structured Extraction: Extracts CDM-compliant data
  3. Policy Evaluation: Evaluates against policy rules
  4. Validation: Validates extracted data
  5. Result Display: Shows extracted data with edit capability

​
3. Review and Save

  1. Review Data: Review extracted CDM data
  2. Edit if Needed: Use clause editor or CDM accordion editor
  3. Save to Library: Save document to library
  4. Broadcast FDC3: Broadcast to desktop applications
  5. Generate Documents: Option to generate LMA templates

​
API Endpoints

​
Extract Document

Extract structured data from document text. Request Body: 
{
  "text": "Credit Agreement text...",
  "filename": "agreement.pdf"
}
Response: Extracted CDM data with policy evaluation results Code Reference: app/api/routes.py (extract endpoint)

​
Upload Document

Upload and extract from document file. Request: Multipart form data with file Response: Document record with extracted CDM data

​
Extraction Chain

​
Simple Extraction Chain

For documents under 50k characters: Code Reference: app/chains/extraction_chain.py

​
Map-Reduce Chain

For larger documents:
  1. Map Phase: Split document into chunks
  2. Extract Phase: Extract from each chunk
  3. Reduce Phase: Combine extractions into final result
Code Reference: app/chains/map_reduce_chain.py

​
CDM Data Structure

Extracted data follows CDM structure: 
{
  "deal_id": "DEAL_2024_001",
  "parties": [
    {
      "name": "Borrower Corp",
      "role": "Borrower",
      "lei": "12345678901234567890"
    }
  ],
  "facilities": [
    {
      "facility_type": "TermLoan",
      "commitment_amount": {
        "amount": 1000000.00,
        "currency": "USD"
      }
    }
  ]
}
Code Reference: app/models/cdm.py (CreditAgreement model)

​
Policy Integration

All extractions are evaluated against policy rules:
  • Sanctions Screening: Check for sanctioned parties
  • Credit Risk: Evaluate creditworthiness
  • ESG Compliance: Check ESG requirements
  • Regulatory Filings: Identify filing requirements
Code Reference: app/services/policy_service.py

​
User Interface

​
Document Parser

Location: client/src/apps/docu-digitizer/DocumentParser.tsx Features:
  • Multimodal Input Tabs: PDF, text, audio, image input
  • Extraction Results: Display extracted CDM data
  • Clause Editor: Edit individual clauses
  • CDM Accordion Editor: Edit CDM structure
  • FDC3 Broadcast: Broadcast to desktop apps
  • Save to Library: Save extracted documents
Access: Navigate to “Document Parser” in top navigation

​
Best Practices

  1. Quality Input: Ensure document text is clear and complete
  2. Review Results: Always review extracted data for accuracy
  3. Edit as Needed: Use editors to correct any errors
  4. Policy Compliance: Verify policy evaluation results
  5. Save Documents: Save to library for future reference

​
Additional Resources

Last Updated: 2026-01-14
Code Reference: client/src/apps/docu-digitizer/DocumentParser.tsx, app/chains/extraction_chain.py, app/api/routes.py