README.md

# CCPA Compliance Analyzer — OPEN HACK 2026

## Solution Overview

This system analyzes natural-language business practice descriptions and determines whether they violate the California Consumer Privacy Act (CCPA). It uses a **RAG (Retrieval-Augmented Generation)** architecture:

1. **Knowledge Base**: The full CCPA statute (key sections 1798.100–1798.135) is pre-encoded as structured text in `ccpa_knowledge.py`, covering all major consumer rights and business obligations.
2. **LLM Inference**: [Llama 3.2 3B](https://ollama.com/library/llama3.2) (via Ollama) receives a system prompt containing the CCPA statute context + the user's business practice prompt and returns a JSON classification.
3. **Rule-Based Fallback**: A deterministic keyword/pattern matcher provides a reliable backup if the LLM is unavailable or returns unparseable output.
4. **FastAPI Server**: Exposes `GET /health` and `POST /analyze` endpoints on port 8000.

**Pipeline**: `POST /analyze` → LLM (Llama 3.2 3B via Ollama) with CCPA context → JSON parse → logic validation → `{"harmful": bool, "articles": [...]}`

---

## Docker Run Command

```bash
docker run --gpus all -p 8000:8000 -e HF_TOKEN=xxx yourusername/ccpa-compliance:latest
```

Without GPU (CPU-only mode, slower):
```bash
docker run -p 8000:8000 yourusername/ccpa-compliance:latest
```

---

## Environment Variables

| Variable | Required | Description |
|---|---|---|
| `HF_TOKEN` | No | Hugging Face access token (not needed for llama3.2 via Ollama) |
| `MODEL_NAME` | No | Ollama model to use (default: `llama3.2:3b`) |
| `OLLAMA_HOST` | No | Ollama server URL (default: `http://localhost:11434`) |

---

## GPU Requirements

- **Recommended**: NVIDIA GPU with ≥4GB VRAM (RTX 3060 or better)
- **CPU-only fallback**: Supported, but inference will be significantly slower (~30-60s per request). The 120s timeout in the test script provides sufficient buffer.
- **Model size**: llama3.2:3b is ~2GB on disk, ~2GB VRAM when loaded

---

## Local Setup Instructions (Fallback — no Docker)

> Use only if Docker fails. Manual deployment incurs a score penalty.

**Requirements**: Linux, Python 3.11+, [Ollama](https://ollama.com)

```bash
# 1. Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# 2. Start Ollama and pull model
ollama serve &
ollama pull llama3.2:3b

# 3. Install Python dependencies
pip install fastapi uvicorn httpx pydantic

# 4. Run the FastAPI server
cd /path/to/ccpa_project
uvicorn app:app --host 0.0.0.0 --port 8000

# 5. Verify it's running
curl http://localhost:8000/health
```

---

## API Usage Examples

### Health Check
```bash
curl http://localhost:8000/health
# Response: {"status": "ok"}
```

### Analyze — Violation Detected
```bash
curl -X POST http://localhost:8000/analyze \
  -H "Content-Type: application/json" \
  -d '{"prompt": "We are selling our customers personal information to data brokers without giving them a chance to opt out."}'

# Response:
# {"harmful": true, "articles": ["Section 1798.120", "Section 1798.100"]}
```

### Analyze — No Violation
```bash
curl -X POST http://localhost:8000/analyze \
  -H "Content-Type: application/json" \
  -d '{"prompt": "We provide a clear privacy policy and allow customers to opt out of data selling at any time."}'

# Response:
# {"harmful": false, "articles": []}
```

### Using docker-compose (with organizer test script)
```bash
docker compose up -d
python validate_format.py
docker compose down
```

---

## Project Structure

```
ccpa_project/
├── app.py              # FastAPI server + LLM/rule-based analysis
├── ccpa_knowledge.py   # CCPA statute knowledge base (RAG source)
├── requirements.txt    # Python dependencies
├── Dockerfile          # Container definition (pre-pulls llama3.2:3b)
├── start.sh            # Container startup (starts Ollama + uvicorn)
├── docker-compose.yml  # Compose config for easy orchestration
└── README.md           # This file
```

---

## Notes on Accuracy

- The system cites sections based on CCPA statute analysis, not keyword matching alone.
- The LLM is instructed to identify **all** violated sections, not just the most obvious one.
- The rule-based fallback provides reliable detection for common violation patterns.
- Incorrect article citations result in zero marks per the scoring rubric, so the system is conservative: it only cites a section when there is clear evidence of a violation matching that section's specific requirements.