Financial Document Update Pipeline

Overview

This system automatically updates financial documents by matching and replacing numerical values between Excel spreadsheets and Word documents. It uses exact text matching and LLM-based verification to ensure accurate updates while maintaining document structure.

Architecture

Core Components

1. Document Processing

   • Excel Parser (funnels/excel_utils.py): Maps cells to headers
   • Word Parser (funnels/extract.py): Extracts sentences with context
   • Processor (utils/document_processing/processor.py): Orchestrates matching and updates

2. LLM Integration

   • Multiple Provider Support:
     • Claude (Anthropic)
     • Qwen (Default)
   • Configurable timeout and model settings
   • Handles text analysis and verification

3. Web Interface

   • Flask-based web application
   • File upload/download functionality
   • Real-time processing status
   • Results visualization with confidence scores

Pipeline

1. Data Loading

   • Excel Parser: Maps each cell to its row and column headers
   • Document Parser: Extracts sentences from Word documents

2. Matching Process

   • Finds exact text matches between Excel headers and document sentences
   • No assumptions made about text content or structure

3. Value Updates

   • Uses LLM to update matched sentences
   • Preserves original sentence structure and formatting
   • Only updates numerical values

4. Confidence Scoring Base confidence of 0.7 for exact matches, with additional factors:

   • Length ratio (0-0.15): Longer matches relative to sentence length
   • Position score (0-0.15): Earlier matches in sentence weighted higher
   • Update success (0-0.1): Ratio of successfully updated values
   • Multiple match penalty: -0.1 per additional match

Usage

Installation

pip install -r requirements.txt

Configuration

Set up environment variables:

LLM_PROVIDER=qwen # or claude
QWEN_API_KEY=sk-xxx  # Your API key
CLAUDE_API_KEY=sk-xxx  # Your API key

Running

python main.py <docx_file> <excel_file>

Or use the web interface:

python app.py  # Starts server on port 8080

Web Interface

python app.py # Starts server on port 8080

REST API

python api.py # Starts server on port 8080

API Endpoints

Document Processing

POST /run-syncspace/
Files:
  - docx_file: Word document (.docx)
  - excel_file: Excel spreadsheet (.xlsx)

Document Download

GET /download
Response: Updated .docx file

Components

Core Processing

• funnels/excel_utils.py: Excel file parsing
• funnels/extract.py: Word document text extraction
• utils/document_processing/processor.py: Main processing logic

Web Interface

• app.py: Flask web application
• templates/index.html: Upload interface

Configuration

• config/: Configuration management
• utils/llm.py: LLM integration
• utils/logging.py: Logging setup

Error Handling

• Size limit: 16MB per file
• Supported formats: .docx, .xlsx only
• Automatic cleanup of temporary files
• LLM timeout: 30 seconds default

Evaluation

Use eval/evaluate_accuracy.py to measure:

• Precision: Correct updates / Total updates
• Recall: Found changes / Actual changes
• F1 Score: Harmonic mean of precision and recall

Project Structure

• api.py: FastAPI server for REST API
• app.py: Flask server for web interface
• main.py: Main processing logic
• utils/: Utility functions and configuration
• templates/: HTML templates for web interface
• static/: Static files for web interface