Overview
Built a RAG pipeline using FAISS and LLMs to classify and assess risks in legal clauses from 1,500+ contracts, achieving 86% classification accuracy and a 0.72 ROUGE-L score.
Problem
Legal teams spend hundreds of hours manually reviewing contracts for risk clauses. The goal was to automate clause-level risk classification and generate plain-English risk summaries using LLMs.
Approach
- Legal-BERT embeddings for domain-aware clause representation
- FAISS vector index for fast similarity retrieval across 1,500+ contracts
- LLaMA-3 zero-shot reasoning across 300+ high-risk clauses
- 40% reduction in retrieval latency through optimized embedding pipeline
Results
| Metric | Score | |--------|-------| | Classification accuracy | 86% | | ROUGE-L score | 0.72 | | Contracts processed | 1,500+ | | Latency reduction | 40% |
Tech Stack
Python LLaMA-3 Legal-BERT FAISS RAG LlamaIndex