Document Understanding + LLM Dynamic Extraction

A Streamlit application that combines Oracle Document Understanding (OCR) with Large Language Models to automatically extract all fields from documents without requiring predefined field specifications.

Features

• Dynamic Field Extraction: Automatically identifies and extracts all key-value pairs from documents
• Multi-language Support: Handles documents in English, Korean, Japanese, Chinese, Arabic, German, and more
• Smart Line Item Grouping: Properly structures invoice line items in tables
• Document Classification: Automatically classifies documents (Invoice, Receipt, Contract, etc.)
• Confidence Scoring: Provides extraction confidence levels

Prerequisites

• Python 3.8+
• Oracle Cloud Infrastructure (OCI) account with Document Understanding service access
• OCI Generative AI service access

Setup

1. Clone the repository

   git clone <repository-url>
   cd du-to-text-llm

2. Create virtual environment

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install dependencies

   pip install -r requirements.txt

4. Configure OCI credentials

   • Set up your OCI config file at ~/.oci/config
   • Update config.py with your:
     • COMPARTMENT_ID
     • OCI model IDs for LLMs

Usage

Run the Streamlit application:

streamlit run pages/run_DU_plus_LLM.py

Then:

1. Upload a PDF or image document
2. Select OCR language (Auto-detect recommended)
3. Choose LLM provider (Meta Llama 3.3 or Cohere Command R)
4. View extracted fields organized by category
5. Optionally save results to JSON

Output

Extracted data is saved to outputs/ directory as JSON files containing:

• Document classification
• All extracted key-value pairs
• Line items (for invoices/receipts)
• Confidence scores
• OCR raw data (optional)

Project Structure

├── pages/
│   └── 01_DU_plus_LLM_dynamic.py  # Main Streamlit app
├── config.py                       # Configuration settings
├── requirements.txt                # Python dependencies
├── .gitignore                      # Git ignore rules
└── README.md                       # This file

License

Proprietary