ClauseEase AI

ClauseEase AI is a local, privacy-preserving document assistant developed as part of the Infosys Springboard Virtual Internship. It processes PDF and text files, extracts and chunks content, detects language, and performs summarization, translation, and document-based question answering all running fully offline. The system uses a custom Streamlit UI and the LLaMA 3 model served locally through Ollama, ensuring that no data leaves the user’s device and enabling secure handling of sensitive legal documents.

The assistant supports:

• High-speed PDF extraction with PyMuPDF
• Automatic language detection using langdetect
• Chunking of long documents for effective context handling
• Real-time token streaming from a locally hosted LLM
• Document-grounded summarization, translation, and Q&A

1. Overview

This project was built to demonstrate how offline LLMs can simplify and interpret legal documents without relying on cloud-based APIs.By running LLaMA 3 through Ollama, the application avoids external API calls entirely.

The Streamlit interface includes multi-file upload support, instant previews, and an interactive chat that streams responses incrementally, offering a smooth LLM-style user experience.

2. Features

PDF Processing

• Text extraction is performed using PyMuPDF (fitz), which provides significantly faster and cleaner results than standard PDF libraries.
• A short preview (first ~500 characters) is shown immediately in the chat interface before full extraction completes.

TXT Processing

• Text files are decoded using UTF-8, with a fallback to Latin-1 for broader compatibility.
• Extraction is nearly instantaneous due to the absence of complex structure.

Language Detection

• Detected from the first 500 characters using langdetect, enabling translation and customized responses.

Summarization, Translation, and Document Q&A

• Summarizes up to 4000 characters into a structured overview.
• Translates the detected language into natural English.
• Answers queries using only the uploaded document, enforcing grounding to reduce hallucinations.

3. Model and Local Inference

The assistant uses:

• Model: llama3:latest
• Runtime: Ollama's local inference server
• Endpoint: http://localhost:11434/api/generate

Responses are streamed token-by-token using iter_lines(), allowing the UI to update incrementally. The interface displays partial output with a cursor indicator and includes a stop button to interrupt generation.

4. Streamlit Interface

The UI includes:

• Gradient-styled layout with custom CSS
• Chat bubbles for user and assistant messages
• Multi-file upload panel
• Real-time streaming display
• Automatic chat history management
• Per-file storage of full text, chunks, language metadata, and original content

Uploaded files appear in the chat with a preview and confirmation message once fully processed.

Instant PDF preview extracted using PyMuPDF

Document translation generated using LLaMA 3

5. Folder Structure

ClauseEase-AI/
│
├── app.py # main streamlit application
├── requirements.txt # dependencies
│
├── screenshots/ # UI images for README
│
└── README.md

6. Running the Project

1. Install Ollama

Download from:
https://ollama.com

2. Pull the model

ollama pull llama3:latest

3. Install Python dependencies

pip install -r requirements.txt

4. Start the application

streamlit run app.py