LLM Guardrails Server

A FastAPI-based moderation server that detects personally identifiable information (PII) in text, optimized for context and compatible with OpenAI's moderation API format.

Features

• ** PII Detection**: Detects various types of personally identifiable information specific to y:

  • phone numbers (landline and mobile)
  • Email addresses
  • Credit card numbers (Visa, MasterCard, AmEx)
  • date formats (DD.MM.YYYY)
  • IP addresses (IPv4 and IPv6)
  • IBAN and BIC codes
  • Medical information with terms

• Dual Detection Engine:

  • Regex-based: Fast pattern matching for structured PII data
  • BERT-based: ML-powered detection for contextual content (optional)

• OpenAI API Compatibility: Drop-in replacement for OpenAI's moderation API

Installation

1. Clone the repository:

git clone <repository-url>
cd llm-guardrails-server

2. Install dependencies:

pip install -r requirements.txt

3. (Optional) For BERT-based detection, ensure PyTorch and transformers are properly installed:

pip install torch transformers

Quick Start

1. Start the server:

python main.py

2. Test with curl (OpenAI-compatible format):

curl http://localhost:8000/v1/moderations \
  -X POST \
  -H "Content-Type: application/json" \
  -d '{
    "model": "moderation-latest",
    "input": "My phone is 01542345678 and E-Mail max@example.de"
  }'

3. Or use the test client:

python test_client.py

API Endpoints

POST /v1/moderations

Analyze text for PII content. Compatible with OpenAI's moderation API.

Request:

{
  "model": "moderation-latest",
  "input": "Text to analyze for  PII"
}

Response:

{
  "id": "modr-xxxxx",
  "model": "moderation-latest",
  "results": [
    {
      "flagged": true,
      "categories": {
        "pii/phone": true,
        "pii/email": false,
        // ... other categories
      },
      "category_scores": {
        "pii/phone": 0.9,
        "pii/email": 0.0,
        // ... other scores
      },
      "category_applied_input_types": {
        "pii/phone": ["text"],
        // ... other applied types
      }
    }
  ]
}

GET /health

Health check endpoint.

GET /v1/models

List available models (for compatibility).

Configuration

Environment variables:

• HOST: Server host (default: 0.0.0.0)
• PORT: Server port (default: 8000)
• USE_BERT: Enable BERT-based detection (default: true)
• FLAGGING_THRESHOLD: Threshold for flagging content (default: 0.5)
• BERT_MODEL_NAME: BERT model to use (default: unitary/toxic-bert)

PII Categories Detected

Category	Description	Examples
pii/phone	phone numbers	030-12345678, +49 30 12345678, 0172-9876543
pii/email	Email addresses	user@domain.de
pii/credit_card	Credit card numbers	4532-1234-5678-9012
pii/ip_address	IP addresses	192.168.1.1
pii/iban	IBAN code	DE89 3704 0044 0532 0130 00

Development

The project structure:

├── main.py                 # FastAPI application
├── models.py              # Pydantic models
├── service.py             # Moderation service logic
├── checkers/              # Detection modules
│   ├── __init__.py
│   ├── regex_checker.py   # Regex-based  PII detection
│   └── bert_checker.py    # BERT-based detection
├── test_client.py         # Test client with  examples
└── requirements.txt       # Dependencies