enabler: guardrails component design

[akotyla]

We should think about introducing a guardrails component to manage multiple validation checks within the system. We should focus on user input.

All checkers should be stackable and run asynchronously.

Guardrails can include for example: safety checker to filter harmful or inappropriate content and relevance checker to verify if the user’s question aligns with the system's purpose and scope.

https://platform.openai.com/docs/models/moderation
https://github.com/unitaryai/detoxify

[micpst]

Resources that can be helpful:

• https://www.llama.com/docs/model-cards-and-prompt-formats/llama-guard-3
• https://www.lakera.ai/lakera-guard

[micpst]

Maybe we should also consider using an LLM proxy for this - https://docs.litellm.ai/docs/simple_proxy