Add `hallucination` check

[linear]

Summary

Add a built-in LLM-based check that detects fabricated facts, invented details, or false claims in model outputs that are not supported by any provided context.

Motivation

Hallucination detection is a core AI safety concern. DeepEval ships HallucinationMetric and Patronus AI built a specialized model (Lynx) that beats GPT-4 on hallucination detection. This is distinct from Groundedness — Groundedness checks if claims are grounded in context, while Hallucination specifically detects fabricated information.

Implementation Guide

Steps

1. Create template: src/giskard/checks/prompts/judges/hallucination.j2
   • Given an answer and optional context, detect fabricated facts
   • Identify specific hallucinated claims with explanations
   • Consider: invented statistics, fake citations, non-existent entities, fabricated details
2. Create check: src/giskard/checks/judges/hallucination.py
   • Subclass BaseLLMCheck, register as "hallucination"
   • Support:
     • answer_key: JSONPathStr — JSONPath for answer (default: trace.last.outputs)
     • context: str | list[str] | None = None — reference context
     • context_key: JSONPathStr | None = None — JSONPath for context
3. Add tests

Distinction from Groundedness

• Groundedness: "Is the answer supported by the context?" — checks grounding
• Hallucination: "Does the answer contain fabricated information?" — checks fabrication, can work with or without context

Example usage

from giskard.checks import Hallucination, Scenario

scenario = (
    Scenario(name="no_hallucination")
    .interact(
        inputs="What year was Python created?",
        outputs="Python was created in 1991 by Guido van Rossum.",
        metadata={"context": "Python was first released in 1991."}
    )
    .check(Hallucination(context_key="trace.last.metadata.context"))
)

Related issues

• Bump @vue/cli-plugin-unit-jest from 4.5.15 to 5.0.3 in /frontend #22-24 (RAG metrics), Bug/design partner ai inspect bugs #33 (RAGET v3)

Acceptance Criteria

• Detects fabricated facts and invented details
• Works with and without provided context
• Identifies specific hallucinated claims in the reason
• Distinct behavior from Groundedness
• Tests cover: factual answer passes, hallucinated answer fails, no context mode

[linear]

OSS-53

[JohnnyTarrr]

We run claim-level hallucination detection in production — VeroQ Shield extracts verifiable claims from any LLM output, checks each against evidence, and returns per-claim verdicts with confidence scores.

The key difference from self-consistency approaches: Shield verifies against external evidence, not the same model. Returns supported/contradicted/unverifiable per claim, with corrections for contradicted claims.

pip install veroq then from veroq import shield; result = shield(text)

Could plug in as an external verification backend for Giskard checks. Happy to help with the interface if useful.

https://github.com/veroq-ai/shield