Contract Analyzer

AI-powered legal contract analysis tool. Upload documents, ask questions, get verified answers with exact quotes from the text.

Features

Part A: Core Features

• Document Upload: Accept PDF and DOCX files with text extraction
• Chat Interface: Ask questions about uploaded contracts
• Quote Verification: All answers backed by verified quotes from the document
• Streaming Responses: Answers appear token-by-token as they're generated
• Large Document Support: Handles 150+ page contracts efficiently via chunking

Part B: Advanced Features

• Citation Highlighting: Click a quote to jump to and highlight the passage in the PDF viewer
• Multi-Document Q&A: Ask comparative questions across multiple documents at once
• Document Comparison: Upload two versions and get a semantic diff showing what changed substantively
• Significance Scoring: Changes ranked by legal/financial impact (1-10 scale) with filtering

Part C: Agentic Research

• Autonomous Tool-Calling: The AI uses tools to research documents independently
  • search_document(query): Search for matching passages
  • get_section(number): Retrieve a specific section
  • list_clauses(): Identify standard contract clauses
• Multi-Round Reasoning: Up to 5 rounds of tool calls per question
• Live Progress Display: Watch as the AI searches and analyzes

Tech Stack

• Frontend: Next.js 15, React, TypeScript, Tailwind CSS
• Backend: Next.js API Routes
• PDF: pdfjs-dist (rendering + text extraction)
• DOCX: mammoth (text extraction)
• LLM: Groq (Mixtral 8x7B) - free tier
• Storage: Filesystem-based with JSON metadata

Getting Started

Prerequisites

• Node.js 18+
• Groq API key (free: groq.com)

Installation

npm install

Configuration

Create .env.local:

GROQ_API_KEY=your_groq_api_key

Run Locally

npm run dev
# Open http://localhost:3000

Build

npm run build
npm start

Usage

1. Standard Chat

• Upload → Open document → Ask question → Get answer with verified quotes
• Click quote to highlight in PDF viewer

2. Multi-Document Q&A

• Select 2+ docs → "Q&A: Multiple Docs" → Ask comparative question
• Quotes labeled with source document

3. Document Comparison

• Select 2 docs → "Compare: Two Versions" → See semantic diff
• Filter by significance level (what changed matters, not just rewordings)

4. Agentic Research

• Open document → "🔬 Agentic Research" → Ask complex question
• Watch AI call tools autonomously, reasoning shown in real-time

Quote Verification

How It Works:

1. LLM returns answer with quotes in [QUOTE]text[/QUOTE] format
2. Backend extracts and normalizes quote (collapses whitespace)
3. Searches for normalized text in document
4. Only shows quote if found (marked "✓ Verified")
5. If not found: marked "⚠ Unverified" or removed

Where It Could Fail:

• Scanned PDF with poor OCR (character errors prevent matching)
• Text extraction artifacts (extra spaces/line breaks)
• Unicode/encoding issues
• LLM heavily paraphrases (by design—we catch this)

Large Documents

For docs > 8K tokens:

1. Split into overlapping chunks (8K tokens, 500 token overlap)
2. For each question, retrieve 5 most relevant chunks
3. Send only relevant chunks to LLM
4. Prevents hallucination about unseen sections

API Endpoints

Method	Path	Purpose
POST	/api/upload	Upload document
GET	/api/documents	List documents
DELETE	/api/documents?id=...	Delete document
POST	/api/chat	Chat single document
POST	/api/multi-chat	Chat multiple documents
POST	/api/compare	Compare two documents
POST	/api/agentic-chat	Agentic research
GET	/api/pdf/[id]	Serve PDF

What's Done

✅ Upload PDF + DOCX
✅ Quote verification
✅ Streaming chat with stop button
✅ PDF citation highlighting
✅ Multi-document Q&A
✅ Semantic document comparison
✅ Agentic research with tool-calling
✅ Error handling + loading states
✅ Responsive UI

What's Not Done

❌ Chat history persistence
❌ DOCX citation highlighting
❌ Scanned PDF detection
❌ Arabic/RTL support
❌ Anonymization
❌ Semantic embeddings
❌ Export as PDF/Word

Part C: Agentic Research

Chosen: Agentic tool-calling (not redlining)

Why: Lower risk, cleaner outcome, visible reasoning, better for analysis

How: LLM calls tools autonomously → max 5 rounds → progress shown → final answer with verified quotes

Deployment

Ready for Vercel/Railway/Render

npm run build
# Deploy to Vercel via GitHub integration or CLI

Performance

• First request: ~3-5s (LLM init)
• Streaming starts: ~2s
• Multi-doc comparison: ~5-10s
• PDF rendering: smooth to ~100 pages

Known Issues

• Large PDFs (500+) may timeout
• Scanned PDFs may have extraction errors
• Mobile layout basic (not fully responsive)
• Quote verification fails on tables with complex formatting

Future Work

1. Persist chat history
2. DOCX highlighting
3. Redlining with tracked changes
4. Semantic search (embeddings)
5. Better OCR handling
6. Export with quotes
7. Multi-language support

---

Built with Next.js + Groq LLM