Skip to main content
The Hallucination Removal Engine is IrisAgent’s pre-delivery check that blocks answers whose claims are not grounded in your indexed content. If a statement cannot be supported by retrieved sources, the answer does not go out as a resolved reply. Operators still control which sources are connected, which procedures are enabled, and which quality rules run after delivery.

​
How validation works before delivery

When IrisAgent prepares an answer (chat, ticket note, copilot suggestion, or API response), the engine:
  1. Retrieves candidate articles, tickets, and procedures for the question
  2. Generates a draft answer from that retrieved set
  3. Checks faithfulness so each material claim lines up with the retrieved source bodies
  4. Attaches citations when the answer passes, or refuses / clarifies when it does not
This is the same grounding loop described in Understanding grounded answers in IrisAgent. The Hallucination Removal Engine is the gate that treats unsupported claims as a delivery failure, not as optional polish. IrisAgent also gates model and prompt changes on two eval sets: questions the AI should resolve, and known-unanswerable questions where the correct behavior is to decline. Changes that raise resolution at the cost of fabrication do not ship.

​
What operators should configure

You do not turn the Hallucination Removal Engine on as a separate toggle. You improve its effectiveness by controlling inputs and follow-up QA: A practical starting AutoQA rule under Resolution Quality:
AI agent must not fabricate or hallucinate product features, pricing, or policies

​
What “zero hallucinations” means here

In IrisAgent product language, zero hallucinations means: unsupported claims are not delivered as grounded answers. The system prefers a decline, a clarifying question, or a human handoff over a confident guess. That is different from claiming every conversation ends in a full automated resolution. Automation rate and grounding are separate goals. Grounding wins when they conflict.

​
Limits to keep in mind

  • Stale or wrong articles still pass faithfulness. If the cited article is outdated or answers a different question, the engine can be faithful to the wrong source. Support Analyst labels that wrong article (a retrieval / KB problem), not a generation hallucination.
  • Incomplete answers can still be grounded. An answer may cite the right article and still omit a required step. That is an incomplete defect in Answer Grounding QA.
  • KB gaps look like refusals. If no article contains the answer, IrisAgent should decline. Fix coverage with new articles or Knowledge Suggestions (AutoKB), not by loosening the gate.
  • Always-on grounding audit of every AI answer (beyond on-demand Support Analyst checks and AutoQA rules) is on the product roadmap. Use spot checks and AutoQA today.
  • Human agents can still edit or override suggested text in Agent Assist. Copilot suggestions are starting points; agents remain accountable for what they send.
Questions about grounding or QA setup: contact@irisagent.com.