grounding-guard

name	grounding-guard
version	1.0.0
description	Verify response faithfulness against RAG source documents using NLI entailment and LLM-as-judge
author	Frame.dev
namespace	wunderland
category	security
tags	["guardrails","hallucination","grounding","faithfulness","nli","rag","fact-checking"]
requires_tools	["check_grounding"]
metadata	{"agentos":{"emoji":"🔍"}}

Grounding Guard

A guardrail automatically verifies that your responses are faithful to the source documents retrieved via RAG. Claims in your output are checked against the retrieved sources using NLI entailment detection.

When to Use check_grounding

• Before presenting synthesized answers from multiple RAG sources
• To verify that summarized content faithfully represents the originals
• When combining information from multiple documents

What It Checks

• Supported: claim is entailed by at least one source document
• Contradicted: claim directly contradicts a source document
• Unverifiable: claim cannot be found in any source (potential hallucination)

Interaction with HITL Guardrail Overrides

When HITL is enabled with guardrailOverride: true, the grounding guard participates in the post-approval guardrail pipeline. If a tool action is approved (by human or LLM judge) but produces output containing ungrounded claims, the grounding guard can flag or veto the response before it reaches the user. See the hitl-safety skill for HITL configuration details.

Constraints

• Only runs when RAG sources are available (no sources → no verification)
• NLI model (~40MB) loads lazily on first grounding check
• LLM escalation for ambiguous claims adds ~150-500ms (only when configured)
• Best for factual/informational responses; less useful for creative/opinion content