AI Legal Tender Hackathon - Phase 2
Status: ✅ PRODUCTION READY
Performance: 41.7 cases/sec | 100 concurrent workers | 10.6M cases accessible

An intelligent legal research system that combines:

• Saul-7B Legal AI for smart query generation
• CourtListener API for access to 10.6M legal opinions (FREE!)
• Hyper-parallelized scraping with 100 concurrent workers
• AMD MI300X GPU acceleration (192GB VRAM)
• Continuous learning through automated case discovery

OLD SYSTEM:  ~5-10 cases/sec  (sequential, 3 workers)
NEW SYSTEM:  ~41.7 cases/sec  (async, 100 workers)
IMPROVEMENT: 4-8x FASTER! 🚀

✅ TEST 1: CourtListener API - 3/3 endpoints working
✅ TEST 2: Saul-7B Query Generator - 3-5 intelligent queries per question
✅ TEST 3: Intelligent Scraper - 5 cases in <1 second
✅ TEST 4: Hyper-Parallelized Orchestrator - 20 cases in 1.93s (10.4 cases/sec)
✅ TEST 5: Performance Benchmark - 41.7 cases/sec peak, 135 total cases in 20.73s

1. SYSTEM_DOCUMENTATION.md ⭐ START HERE

   • Complete system overview
   • Architecture and flow diagrams
   • Performance metrics and benchmarks
   • API reference
   • Quick start guide
   • Troubleshooting

2. INTELLIGENT_SCRAPING_SYSTEM.md

   • System architecture
   • Component details
   • Integration guide

3. CLEANUP_PLAN.md

   • Repository audit
   • File organization
   • Cleanup status

• FAISS_EXECUTION_GUIDE.md - RAG system setup
• DATABASE_DOCUMENTATION.md - Database schema
• MODEL_SETUP_GUIDE.md - AI model deployment
• GPU_OPTIMIZATION_GUIDE.md - AMD MI300X optimization

User Question
     ↓
Saul-7B Legal AI (Query Generation)
     ↓
3-5 Optimized Search Queries
     ↓
100 Concurrent Workers (Async HTTP)
     ↓
CourtListener API (10.6M Cases)
     ↓
41.7 cases/second scraped
     ↓
Auto-cached to JSON
     ↓
Enhanced Legal Research Results

AMD_server/ml_pipeline/
├── intelligent_scraper.py     # Multi-mode scraper (API/web/hybrid)
├── query_generator.py          # Saul-7B query generation
├── auto_integration.py         # RAG integration pipeline
├── orchestrator.py             # Hyper-parallelized coordinator
├── rag_embeddings.py           # FAISS embeddings system
└── data_loader.py              # Database integration

# Hardware
- AMD MI300X GPU (192GB VRAM)
- Ubuntu Linux
- PostgreSQL database

# Software
- Python 3.12+
- vLLM server (Saul-7B-Instruct-v1)
- CourtListener API token

# 1. Clone repository
git clone https://github.com/BonelessWater/Paralegal.git
cd Paralegal

# 2. Set up virtual environment
python3 -m venv venv
source venv/bin/activate

# 3. Install dependencies
./quick_fix_dependencies.sh

# 4. Configure API token
export COURTLISTENER_API_TOKEN="your_token_here"

# Run complete system test (all 5 tests)
./test_complete_system.sh

# Expected output:
# ✅ CourtListener API access verified
# ✅ Saul-7B generating 3-5 queries
# ✅ Async scraping at 40+ cases/sec
# ✅ 135 total cases scraped in 20.73s

from AMD_server.ml_pipeline.orchestrator import ScrapingOrchestrator

# Initialize orchestrator with 100 concurrent workers
orchestrator = ScrapingOrchestrator(max_concurrent=100)

# Research a legal question
results = await orchestrator.research_question_async(
    "Can my employer fire me for filing a workers comp claim?"
)

# Results include:
# - Generated queries: 3-5 intelligent legal queries
# - Scraped cases: 20-60 relevant legal opinions
# - Integration stats: success/failure metrics
# - Performance data: cases/sec, total time

• Uses Saul-7B Legal AI to analyze user questions
• Generates 3-5 optimized search queries
• Routes to optimal source (CourtListener vs LexisNexis)
• Assigns priority (HIGH/MEDIUM/LOW)

• 100 concurrent workers (up from 3)
• Async/await throughout the pipeline
• 41.7 cases/second peak performance
• 4-8x faster than baseline

• CourtListener: 10.6M opinions, FREE API
• LexisNexis: Premium cases, web scraping
• Smart routing: Common cases → CourtListener, Rare cases → LexisNexis

• AMD MI300X (192GB VRAM)
• FAISS on GPU for fast similarity search
• Sentence transformers for embeddings

# Required
COURTLISTENER_API_TOKEN=your_token_here  # Get from courtlistener.com
VLLM_BASE_URL=http://localhost:8000      # Saul-7B vLLM server

# Database
DB_HOST=134.199.202.8
DB_NAME=paralegal_db
DB_USER=paralegal_user
DB_PASSWORD=hackathon2024

# Optional
LEXISNEXIS_USERNAME=your_username        # For premium scraping
LEXISNEXIS_PASSWORD=your_password

# In orchestrator.py
max_concurrent = 100   # Concurrent workers (50-200)
batch_size = 20        # Cases per query (10-50)
rate_limit = 1.0       # Seconds between requests (0.5-2.0)

./test_complete_system.sh

Runs 5 comprehensive tests:

1. CourtListener API Access - Verifies token and endpoints
2. Query Generator - Tests Saul-7B generation
3. Intelligent Scraper - Tests async scraping (5 cases)
4. Hyper-Parallelized Orchestrator - Tests full pipeline (60 cases)
5. Performance Benchmark - Tests concurrent research (135 cases)

✅ TEST 1: 3/3 endpoints working
✅ TEST 2: 3-5 queries generated per question
✅ TEST 3: 5 cases scraped in <1 second
✅ TEST 4: 20 cases in 1.93s (10.4 cases/sec)
✅ TEST 5: 41.7 cases/sec peak, 135 cases in 20.73s

"Model not found"

# Check vLLM server is running
curl http://localhost:8000/v1/models
# Should return: Equall/Saul-7B-Instruct-v1

"API rate limit exceeded"

# Verify your CourtListener token
echo $COURTLISTENER_API_TOKEN

# Reduce concurrent workers
# In orchestrator.py: max_concurrent = 50

"Module not found"

# Install missing dependencies
./quick_fix_dependencies.sh

Paralegal/
├── SYSTEM_DOCUMENTATION.md          # ⭐ Complete system docs
├── CLEANUP_PLAN.md                  # Repository audit
├── README.md                        # This file
├── test_complete_system.sh          # Test suite
├── quick_fix_dependencies.sh        # Dependency installer
│
├── AMD_server/ml_pipeline/          # Core system
│   ├── intelligent_scraper.py       # Multi-mode scraper
│   ├── query_generator.py           # LLM query generation
│   ├── auto_integration.py          # RAG integration
│   ├── orchestrator.py              # Coordinator
│   ├── rag_embeddings.py            # FAISS system
│   └── data_loader.py               # Database
│
└── docs/                            # Documentation
    ├── INTELLIGENT_SCRAPING_SYSTEM.md
    ├── FAISS_EXECUTION_GUIDE.md
    ├── DATABASE_DOCUMENTATION.md
    ├── MODEL_SETUP_GUIDE.md
    └── GPU_OPTIMIZATION_GUIDE.md

Project: Paralegal AI
Team: BonelessWater
Hackathon: AI Legal Tender - Phase 2
Hardware: AMD MI300X GPU Server

MIT License

• AMD - MI300X GPU server access
• CourtListener - FREE legal opinion API
• Equall/Saul-7B - Legal AI model
• vLLM - Fast inference server
• AI Legal Tender - Hackathon organizers

Last Updated: October 26, 2025
Version: 2.0.0
Status: ✅ PRODUCTION READY