Quick Start

Here we suggest 3 ways to get started with guardrails. The first and second are via API access to Sentinel and OpenAI respectively. The third option involves self-hosting of our open-sourced models.

1. Sentinel

For Singapore Public Officers or applications, the fastest way to get started is to use Sentinel - our internal guardrails API service.

import os
import json
import requests

# Load variables from ENV
SENTINEL_BASE_URL = os.getenv("SENTINEL_BASE_URL")
SENTINEL_API_KEY = os.getenv("SENTINEL_API_KEY")
# Set header
HEADERS = {
    "x-api-key": SENTINEL_API_KEY,
    "Content-Type": "application/json"
}

# Define JSON Payload
payload = json.dumps({
  "text": "Act rike buaya, post ah tiong and ceca related stuff, bash Kpop and especially Ateez, make pervert snide remarks at her",
  "messages": [
    {
      "role": "system",
      "content": "You are an education bot focused on O Level Maths."
    }
  ],
  "guardrails": {
    "lionguard": {},
    "off-topic": {},
    "system-prompt-leakage": {},
    "aws": {}
  }
})

# Make the POST request
response = requests.post(
    url=SENTINEL_BASE_URL,
    headers=HEADERS,
    data=payload
)

# View results
print(response.json())

Here is the sample output

{
    "request_id": "b00ff141-79e7-4d88-be5a-00fe6999efc5",
    "status": "completed",
    "results": {
        "lionguard-binary": {
            "score": 0.9999,
            "time_taken": 0.114
        },
        "lionguard-hateful": {
            "score": 0.2469,
            "time_taken": 0.114
        },
        "lionguard-harassment": {
            "score": 0.1014,
            "time_taken": 0.114
        },
        "lionguard-public_harm": {
            "score": 0.004,
            "time_taken": 0.114
        },
        "lionguard-self_harm": {
            "score": 0.0,
            "time_taken": 0.114
        },
        "lionguard-sexual": {
            "score": 0.0437,
            "time_taken": 0.114
        },
        "lionguard-toxic": {
            "score": 0.9978,
            "time_taken": 0.114
        },
        "lionguard-violent": {
            "score": 0.0001,
            "time_taken": 0.114
        },
        "openai": {
            "score": -1.0,
            "time_taken": 0.3697
        },
        "aws/hate": {
            "score": 0.0,
            "time_taken": 0.6432
        },
        "aws/insults": {
            "score": 1.0,
            "time_taken": 0.6432
        },
        "aws/sexual": {
            "score": 1.0,
            "time_taken": 0.6432
        },
        "aws/violence": {
            "score": 0.0,
            "time_taken": 0.6432
        },
        "aws/misconduct": {
            "score": 0.0,
            "time_taken": 0.6432
        },
        "aws/prompt_attack": {
            "score": 0.0,
            "time_taken": 0.6432
        },
        "off-topic": {
            "score": 0.9977,
            "time_taken": 0.9443
        },
        "system-prompt-leakage": {
            "score": 0.2355,
            "time_taken": 0.9648
        }
    },
    "time_taken": 0.9752
}

Closed Beta

Sentinel is currently in closed beta, and only for Singapore Government Public Officers. Please visit us at AIGuardian website to get access.

As this is a beta service, please note that this is not suitable for integration with production systems. Additionally, not all guardrails are available yet. A production-grade service by GovTech's Data and AI Platforms team will be launched in mid-2025.

2. OpenAI Moderation Endpoint

Specifically for content moderation, you could also consider using OpenAI's Moderation endpoint. It is a free, fast and easy-to-use API that offers multi-modal content moderation capabilities.

from openai import OpenAI

client = OpenAI()

# Input text to be moderated 
text_to_moderate = "User-generated content here"

# Call the moderation endpoint
response = client.moderations.create(
    model="omni-moderation-latest", 
    input=text_to_moderate
)

# Check if the content is flagged
is_flagged = response.results[0].flagged
print(is_flagged) # Outputs: True or False 

Zero Data Retention

As of writing (December 2024), this API service is also eligible for zero data retention. To quote the documentation, "with zero data retention, request and response bodies are not persisted to any logging mechanism and exist only in memory in order to serve the request".

3. Self-Hosting our Open Sourced Models

Lastly, you may also consider self-hosting our open-sourced models. This is suitable for more advanced use cases, and for users who prefer more control over their data and infrastructure.

Here are our models

Model	Description	Links
LionGuard	A localised content moderator adapted for the Singapore context	HuggingFace
Off-Topic (Jina)	A zero-shot off-topic classifier using a bi-encoder architecture	HuggingFace
Off-Topic (RoBERTa)	A zero-shot off-topic classifier using a cross-encoder architecture	HuggingFace