PrivyDoc - Local Document Intelligence Tool

A secure, on-device document analysis solution powered by Foundry Local, designed to handle sensitive documents without relying on the cloud.

---

Overview

The PrivyDoc Document Analysis Tool transforms how teams work with sensitive documents by bringing AI-powered analysis directly to your device. Using Microsoft Foundry Local, all processing happens locally-no data leaves your computer.

Ideal for:

• Policy teams analyzing internal documents
• Researchers handling sensitive research materials
• Legal teams reviewing confidential contracts
• Compliance teams processing regulatory documents
• Anyone needing document insights without uploading to the cloud

Key Features

Document Processing

• ✅ Multi-Format Support: PDF and DOCX
• ✅ Structure Recognition: Identify document sections and hierarchy
• ✅ Text Extraction: Preserve formatting cues

AI-Powered Insights

• ✅ Smart Summarization: Concise overviews of entire documents or sections
• ✅ Entity Recognition: Detect people, organizations, locations, dates, etc.
• ✅ Sentiment Analysis: Analyze emotional tone at document and section levels
• ✅ Topic Classification: Auto-categorize documents by subject

Security & Compliance

• ✅ 100% Local Processing
• ✅ Zero Data Transmission
• ✅ Air-Gap Compatible
• ✅ Analysis Traceability: Logs of all interactions
• ✅ Document Fingerprinting: Verify integrity and processing history

Export & Integration

• ✅ Multiple Formats: Markdown, JSON, CSV
• ✅ Structured Data: Standardized output for downstream processing
• ✅ Analysis History: Browse previous analyses with metadata

Getting Started

System Requirements

Component	Requirement
OS	Windows 10/11, macOS 12+
RAM	8 GB minimum, 16 GB recommended
Storage	≥ 5 GB free
Python	3.10+

Installation

1. Clone the repository

git clone https://github.com/ShivamGoyal03/PrivyDoc.git
cd PrivyDoc

2. Set up Python environment

# Create a virtual environment
python -m venv venv

# Activate it
venv\Scripts\activate   # Windows
source venv/bin/activate  # macOS/Linux

# Install dependencies
pip install -r requirements.txt

3. Install Foundry Local

# Windows
winget install Microsoft.FoundryLocal

# macOS
brew tap microsoft/foundrylocal
brew install foundrylocal

# Verify installation
foundry --version

4. Download a model (automatic on first run)

# List available models
foundry model ls

# Download specific model (optional)
foundry model run qwen2.5-0.5b

Usage

Web Interface (Recommended)

chainlit run multi_agent_doc_analysis.py

• Open: http://localhost:8000
• Upload PDF or DOCX, track progress, export results

Command Line

python multi_agent_doc_analysis.py --file path/to/document.docx

• Process a single file and save results to local JSON database

Supported Models

Model	Size	Best For	Notes
qwen2.5-0.5b	Small	Quick analysis	Default, fastest
phi-3.5-mini	Medium	Balanced performance	Good all-around
phi-4	Large	Detailed analysis	Most accurate

Implementation Details

Architecture

• Document Processor: Extracts and normalizes text
• AI Engine: Runs Foundry Local models
• Analysis Workflow: Multi-step analysis using Agent Framework
• Storage Layer: JSON-based analysis history
• User Interfaces: Web UI (Chainlit) & CLI

Processing Flow

1. Document Upload (via Web or CLI)
2. Text Extraction (PDF/DOCX cleaned for analysis)
3. Section Extraction (LLM-powered agent)
4. Entity Extraction (NER agent identifies people, orgs, locations)
5. Summarization & Sentiment (Analyzer agent)
6. Results Compilation & Storage (saved locally in JSON)
7. UI Feedback/Export (Markdown, JSON, CSV)

Processing Pipeline

flowchart TD
    A([User Uploads PDF/DOCX])
    B([Text Extraction])
    C([Section Extraction Agent])
    D([Entity Recognition Agent])
    E([Summarize + Sentiment Agent])
    F([Results Saved Locally])
    G([Chainlit Web UI Shows Results/Export])
    H((User))

    A --> B
    B --> C
    C --> D
    D --> E
    E --> F
    F --> G
    G -- "Download (Markdown/JSON/CSV)" --> H
    B -. CLI Mode .-> F

Loading

Local Storage

All results are saved in analysis_history.json for:

• Historical reference
• Audit purposes
• Quick retrieval without reprocessing

Development

Project Structure

PrivyDoc/
  ├── multi_agent_doc_analysis.py    # Main application
  ├── analysis_history.json          # Local analysis
  ├── requirements.txt
  └── README.md

Dependencies

• foundrylocal: Microsoft Foundry Local SDK
• chainlit: Web UI framework
• pdfplumber: PDF text extraction
• python-docx: DOCX text extraction
• agent_framework: Create & orchestrate agents

Contributing & License

Contributions are welcome! Fork the repo, create your feature or fix branch, and submit a PR.

Licensed under the MIT License — see LICENSE.

Contact

---

Shivam Goyal