A web application for extracting structured data from documents with AI-powered intelligence and feedback learning. Built on Google's LangExtract and Stanford's DSPy frameworks with an intuitive web interface.

• PDF Upload: Convert PDFs to text using OCR
• Text Input: Process plain text content directly
• Multiple Formats: Handle various document types (contracts, invoices, forms)

• Extract Mode: Find specific text spans from documents (names, dates, amounts)
• Generate Mode: Create interpreted content (summaries, classifications)

• Multiple Providers: Azure OpenAI, OpenAI, Google Gemini, Ollama
• Template System: Save and reuse extraction configurations
• Feedback Learning: Improve results through user feedback and DSPy optimization

• Template Management: Create and edit extraction schemas
• Real-time Processing: Live progress updates for document processing
• Result Export: Download results in HTML, JSON, or JSONL formats

• Python 3.8+
• API access to one of: Azure OpenAI, OpenAI, Google Gemini, or Ollama

1. Clone and install

git clone https://github.com/LM-150A/docflash.git
cd docflash
python -m venv venv
source venv/bin/activate  # Windows: venv\Scripts\activate
pip install -r requirements.txt

2. Configure environment

cp .env.example .env
# Edit .env with your API credentials

3. Run application

python start_fastapi.py

Access the application at http://localhost:5000

# Choose one provider
LLM_PROVIDER=azure_openai  # Options: azure_openai, openai, gemini, ollama

# Azure OpenAI
AZURE_OPENAI_ENDPOINT=https://your-resource.openai.azure.com/
AZURE_OPENAI_DEPLOYMENT_NAME=gpt-4
AZURE_OPENAI_API_KEY=your-api-key

# OpenAI
OPENAI_API_KEY=your-openai-key

# Google Gemini  
GOOGLE_API_KEY=your-google-api-key

# Ollama (local)
OLLAMA_MODEL_ID=gemma2:2b
OLLAMA_BASE_URL=http://localhost:11434

# Optional: Enable DSPy optimization
DSPY_ENABLED=true

Create extraction attributes specifying what information to extract:

• Upload PDF files for OCR processing
• Copy/paste text content directly
• Provide multiple examples for better training

The system creates training examples based on your schema and sample documents.

Upload new documents and run extraction with configurable settings:

• Number of extraction passes (1-3)
• Parallel processing workers (5-20)
• Temperature settings based on extraction modes

• Rate generated examples to improve future results
• Use detailed feedback to guide AI improvements
• DSPy automatically optimizes prompts based on feedback

Frontend (HTML/JS) ←→ Backend (FastAPI) ←→ AI Providers
    ↓                       ↓                    ↓
• Template UI          • LangExtract        • Azure OpenAI
• Document Upload      • OCR Pipeline       • OpenAI  
• Feedback System      • DSPy Integration   • Google Gemini
• Progress Tracking    • Template Storage   • Ollama

• Use Extract mode for factual data that appears verbatim
• Use Generate mode for analysis or interpreted content
• Write clear, specific attribute descriptions

• Provide 2-4 diverse sample documents
• Include variations and edge cases
• Ensure samples cover all schema attributes

• Rate examples regularly to improve performance
• Use detailed feedback for specific issues
• Feedback is isolated by document type

API Configuration

• Verify API credentials in .env file
• Check endpoint URLs and model names
• Ensure sufficient API quota/credits

PDF Processing

• Use clear, text-based PDFs (not scanned images)
• Check file size limits (typically 16MB max)
• Try alternative OCR if text extraction fails

Poor Extraction Results

• Review and improve schema descriptions
• Add more diverse training examples
• Increase extraction passes for better recall
• Provide feedback on generated examples

DSPy Optimization

• Set DSPY_ENABLED=true in environment
• Provide sufficient feedback (default: 10+ examples)
• Check logs for optimization triggers

1. Fork the repository
2. Create a feature branch
3. Make changes and add tests
4. Submit a pull request

Apache License 2.0 - see LICENSE file for details.

• Google LangExtract - Core extraction framework
• Stanford DSPy - Prompt optimization framework