Sentinel
Sentinel is GovTech's internal guardrail API service that provides a comprehensive suite of input and output guardrails for AI applications. It offers protection against risks like toxicity, prompt injection, PII exposure, and hallucination.
Sentinel is designed to be:
- Fast and performant with minimal latency impact
- Model-agnostic and compatible with any LLM
- Highly configurable with adjustable confidence thresholds
- Easy to integrate via REST API
You can explore Sentinel's capabilities through:
- Web demo: 🌐 Try Sentinel
- API documentation: 📚 Sentinel Docs
Currently in closed beta, Sentinel is available exclusively to Singapore Public Officers and government applications. To request access, please contact gabriel_chua@tech.gov.sg.
Available Guardrails
| Name | Description | Remarks |
|---|---|---|
| LionGuard | An input/output guardrail to detect Not Safe For Work (NSFW) language in the Singapore context | - |
| PromptGuard | An input guardrail to detect prompt injection and jailbreak attempts | - |
| Off-Topic | An input guardrail to detect off-topic responses | - |
| System-Prompt-Leakage | An output guardrail to detect exposure of substantive information of the system prompt | Coming soon |
| Cloak | An input/output guardrail to detect personally identifiable information (PII), developed by GovTech | Direct integration to cloak.gov.sg coming soon |
Benchmarking
Coming Soon
We will be releasing a benchmarking report soon.