Contract Review System

Overview

Contract Review System is a Retrieval-Augmented Generation (RAG)-based legal document analysis platform designed to automate contract understanding, compliance checking, clause summarization, and risk assessment.

The project was developed as part of my major project focusing on explainable and modular AI-assisted contract analysis.

Features

1. Contract Summarization

• Extracts key contract sections
• Generates concise clause summaries
• Uses retrieval-based contextual summarization

2. Compliance Checking

• Rule-based compliance evaluation
• Semantic retrieval using FAISS
• Multi-stage keyword validation pipeline
• Evidence-based clause extraction

3. Risk Analysis

• Business risk assessment based on compliance results

• Generates:

  • Risk level
  • Business impact
  • Recommendations

4. PDF Report Generation

• Generates structured downloadable PDF reports

• Includes:

  • Clause summaries
  • Compliance results
  • Risk analysis

---

Retrieval Pipeline

The system follows a Retrieval-Augmented Generation (RAG) architecture:

1. PDF text extraction
2. Chunk creation with overlap
3. Embedding generation using MiniLM
4. Vector storage in FAISS
5. Semantic retrieval using rule-based queries
6. LLM-based analysis on retrieved chunks

---

Compliance Checking Workflow

Rule Selection
      ↓
Query Generation
      ↓
FAISS Retrieval
      ↓
Keyword Validation
      ↓
LLM Clause Extraction
      ↓
Evidence Verification
      ↓
COMPLIANT / NON-COMPLIANT

The compliance system uses a hybrid retrieval-validation approach combining semantic retrieval with strict keyword-based verification.

---

Tech Stack

Component	Technology
Frontend	Streamlit
LLM	Fine-Tuned Mistral 7B
Embeddings	all-MiniLM-L6-v2
Vector Database	FAISS
PDF Processing	PyMuPDF
Report Generation	ReportLab
Inference Hosting	Google Colab + Ngrok
Training Framework	Hugging Face + LoRA/QLoRA

---

Project Structure

Contract-Review-System/
│
├── app.py
├── requirements.txt
├── README.md
├── LICENSE
├── .gitignore
├── .env.example
│
└── src/
    │
    ├── compliance_checker.py
    ├── key_clause_summarizer.py
    ├── llm_client.py
    ├── pdf_processor.py
    ├── report_generator.py
    ├── risk_and_reco.py
    └── vector_store.py

---

Excluded Folders

The following folders are excluded from the repository:

Folder	Reason
models/	Large GGUF model files
data/	Generated vector indexes and metadata
logs/	Runtime logs
config/	Environment-specific configuration
trainingV2/	Training datasets and generated training files

---

Training Process

The base Mistral-7B Instruct v0.2 model was fine-tuned for legal contract analysis using LoRA/QLoRA with Unsloth for memory-efficient training. The model was trained on legal-domain instruction datasets to improve contract understanding, summarization, clause interpretation, compliance reasoning, and risk assessment.

Datasets Used

• CUAD (Contract Understanding Atticus Dataset)
• ACORD Dataset
• Custom compliance-rule datasets

Fine-Tuning Configuration

• Parameter-efficient fine-tuning using LoRA/QLoRA
• LoRA adapters applied to the transformer's attention (q_proj, k_proj, v_proj, o_proj) and feed-forward (gate_proj, up_proj, down_proj) layers
• Tuned key hyperparameters including LoRA rank, LoRA alpha, learning rate, batch size, gradient accumulation, and maximum sequence length
• 4-bit quantization for memory-efficient training and inference
• Training performed using Supervised Fine-Tuning (SFT) with the Unsloth framework

Training Objectives

• Legal clause understanding
• Contract summarization
• Compliance-oriented reasoning
• Risk identification and recommendations
• Question answering over legal contracts

Export

The fine-tuned model was exported to GGUF (Q4_K_M) format for efficient inference using llama.cpp, and later served through FastAPI for integration with the RAG pipeline.

Key Research Contributions

• Hybrid RAG-based legal analysis pipeline
• Rule-driven compliance evaluation
• Multi-stage evidence verification
• Modular explainable architecture
• Integration of retrieval and legal reasoning workflows

---

Limitations

• Compliance logic is primarily keyword-guided
• Requires active remote inference server
• Performance depends on retrieval quality
• Does not replace professional legal review

---

Future Improvements

• Full logical compliance reasoning
• Multi-contract comparison
• Clause highlighting in PDFs
• Persistent analysis caching
• Improved legal reasoning datasets
• Multi-agent legal workflows

---

Disclaimer

This project is intended for research and educational purposes only. It does not provide legal advice. It is under MIT License