📄 DocExtract AI

An intelligent agentic document extraction solution built with Streamlit, FastAPI, and LangGraph. Extract structured data from PDF documents using AI, with human-in-the-loop feedback for refinement.

✨ Features

• 🤖 Agentic Extraction: LangGraph-powered workflow for intelligent document understanding
• 📋 Schema-Driven: Define extraction fields via JSON schema
• 👁️ Side-by-Side View: PDF viewer alongside extracted data
• 💬 Human-in-the-Loop: Provide feedback to refine extractions
• 🔄 Iterative Refinement: Re-run extraction with corrections
• 📊 Confidence Scores: See extraction confidence for each field
• 🎨 Modern UI: Beautiful, responsive Streamlit interface

🏗️ Architecture

┌─────────────────────────────────────────────────────────────┐
│                    Streamlit Frontend                        │
│  ┌──────────────────┐      ┌──────────────────────────┐     │
│  │   PDF Viewer     │      │   Extraction Results     │     │
│  │   - Page nav     │      │   - Field values         │     │
│  │   - Zoom         │      │   - Confidence bars      │     │
│  └──────────────────┘      │   - Feedback input       │     │
│                            └──────────────────────────┘     │
└─────────────────────────────┬───────────────────────────────┘
                              │ HTTP/REST
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                    FastAPI Backend                           │
│  ┌────────────┐  ┌────────────┐  ┌──────────────────────┐   │
│  │  Upload    │  │  Extract   │  │     Feedback         │   │
│  │  /upload   │  │  /extract  │  │     /feedback        │   │
│  └────────────┘  └────────────┘  └──────────────────────┘   │
└─────────────────────────────┬───────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                   LangGraph Agent                            │
│  ┌──────────┐    ┌──────────┐    ┌──────────────────────┐   │
│  │ Extract  │───▶│ Validate │───▶│ Apply Feedback       │   │
│  │   Node   │    │   Node   │    │      Node            │   │
│  └──────────┘    └──────────┘    └──────────────────────┘   │
│        │                                    │               │
│        └────────────◀───────────────────────┘               │
│                    (Re-extraction loop)                      │
└─────────────────────────────────────────────────────────────┘

🚀 Quick Start

Prerequisites

• Python 3.9+
• OpenAI API key

Installation

1. Clone or navigate to the project:

   cd doc_extraction

2. Create a virtual environment:

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

3. Install dependencies:

   pip install -r requirements.txt

4. Set up environment variables:

   # Create .env file
   echo "OPENAI_API_KEY=your-openai-api-key-here" > .env

Running the Application

Option 1: Using the startup script

chmod +x start.sh
./start.sh

Option 2: Manual startup

Terminal 1 - Start the API:

python run_api.py

Terminal 2 - Start the frontend:

streamlit run streamlit_app.py

Access the Application

• Frontend: http://localhost:8501
• API Docs: http://localhost:8000/docs

📖 Usage Guide

1. Define Your Schema

In the sidebar, define a JSON schema for the fields you want to extract:

{
  "type": "object",
  "properties": {
    "invoice_number": {
      "type": "string",
      "description": "Invoice or document number"
    },
    "date": {
      "type": "string",
      "description": "Document date"
    },
    "total_amount": {
      "type": "number",
      "description": "Total amount"
    }
  },
  "required": ["invoice_number", "date"]
}

2. Upload a PDF

Drag and drop or click to upload your PDF document.

3. Extract Fields

Click "🚀 Extract Fields" to run the AI extraction.

4. Review Results

• View extracted values alongside their confidence scores
• Fields with low confidence are highlighted
• Navigate PDF pages to verify extraction

5. Provide Feedback (Optional)

If results need correction:

1. Enter feedback in the text area (e.g., "The date should be in MM/DD/YYYY format")
2. Click "🔄 Re-extract with Feedback"
3. The agent will re-process considering your feedback

6. Export Results

Click "📥 Export Results" to download the extracted data as JSON.

🔧 Configuration

Environment Variables

Variable	Description	Default
OPENAI_API_KEY	OpenAI API key (required)	-
OPENAI_MODEL	Model to use	gpt-4o
FASTAPI_HOST	API host	localhost
FASTAPI_PORT	API port	8000

Customizing the Schema

The JSON schema supports:

• String fields: Text values
• Number fields: Numeric values
• Array fields: Lists of items
• Nested objects: Complex structures
• Required fields: Mark fields that must be extracted

📁 Project Structure

doc_extraction/
├── agent/
│   ├── __init__.py
│   └── extraction_agent.py    # LangGraph extraction workflow
├── api/
│   ├── __init__.py
│   ├── main.py               # FastAPI endpoints
│   └── models.py             # Pydantic schemas
├── utils/
│   ├── __init__.py
│   └── pdf_processor.py      # PDF text/image extraction
├── uploads/                   # Uploaded PDFs (auto-created)
├── config.py                 # Configuration settings
├── requirements.txt          # Python dependencies
├── run_api.py               # API startup script
├── start.sh                 # Combined startup script
├── streamlit_app.py         # Streamlit frontend
└── README.md

🛠️ API Reference

Endpoints

Method	Endpoint	Description
POST	/upload	Upload a PDF file
POST	/extract/{session_id}	Extract fields from PDF
POST	/feedback	Submit feedback and re-extract
GET	/pdf/{session_id}/page/{page_num}	Get PDF page as image
GET	/pdf/{session_id}/info	Get PDF metadata
GET	/session/{session_id}	Get session status
DELETE	/session/{session_id}	Delete session

Example API Usage

import httpx
import json

# Upload PDF
with open("document.pdf", "rb") as f:
    response = httpx.post(
        "http://localhost:8000/upload",
        files={"file": f}
    )
    session_id = response.json()["session_id"]

# Extract fields
schema = {
    "type": "object",
    "properties": {
        "title": {"type": "string"},
        "date": {"type": "string"}
    }
}

response = httpx.post(
    f"http://localhost:8000/extract/{session_id}",
    data={"json_schema": json.dumps(schema)}
)
print(response.json())

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

📄 License

This project is licensed under the MIT License.