Skip to main content
Version: Next

RAG Output Guardrails

This page describes the initial output guardrails hook added to RAGEngine chat completions.

Scope​

This document reflects the current non-streaming implementation only.

At this stage, the feature is limited to:

  • assistant text returned by /v1/chat/completions
  • non-streaming responses
  • runtime configuration loaded from environment variables inside the RAG service

It does not yet describe a full CRD-backed end-user configuration flow.

What The Hook Does​

RAGEngine can scan assistant output before returning the final response to the caller.

If configured checks are triggered, the response can be handled in one of two ways:

  • redact: replace matching content with sanitized output
  • block: replace the assistant message with a fixed block message

Current Runtime Configuration​

The current implementation reads these environment variables at runtime:

  • OUTPUT_GUARDRAILS_ENABLED
  • OUTPUT_GUARDRAILS_ACTION_ON_HIT
  • OUTPUT_GUARDRAILS_REGEX_PATTERNS
  • OUTPUT_GUARDRAILS_BANNED_SUBSTRINGS
  • OUTPUT_GUARDRAILS_BLOCK_MESSAGE

These variables are consumed by the RAGEngine process when it starts.

Current Limitations​

  • This page documents the initial hook only.
  • It does not imply that a standard KAITO RAGEngine spec field is already available for end users.
  • Streaming response scanning is not included in this stage.
  • Auditing and richer operational controls are not included in this stage.