Evaluate the options for implementing safety guardrails for locally run LLMs.

• Does not involve fine tuning of models.
• Must be run locally and offline.
• Low latency
  • Less than 0.8 seconds?
• Accuracy
  • False Positives can lead to the app's unusability and missing information.
  • False Negatives can lead to vulnerabilities.
  • If the ideal balance cannot be achieved, the FP-FN tradeoff can be decided based on the LLM use-case.
• Low memory footprint
  • Less that 800 MB?

• Profanity and toxic content
• Prompt Injection
• PII
  • Should certain PII's be allowed? E.g. In a RAG application, extracting email ids of the authors from a whitepaper.

1. Guardrails-AI
2. NeMo-Guardrails
3. Safety-System-Prompts

1. Model-Safety
2. PyRit

1. Use non-LLM, lightweight models from Guardrails-AI to validate the user input and LLM output.
2. Use system prompt engineering similar to Safety-System-Prompts to instruct the LLM itself to behave in a safe manner.