About the Hallucination Removal Engine

How IrisAgent checks AI answers against your content before delivery, what you can configure, and the limits.

The Hallucination Removal Engine blocks answers whose claims are not supported by your indexed content. If a statement cannot be backed by a retrieved source, the answer is not sent as a resolved reply. It is always on. There is no separate toggle.

​

How it works

For every answer (chat, ticket note, copilot suggestion or API response), IrisAgent:

1. Retrieves relevant articles, tickets and procedures.
2. Generates a draft from those sources.
3. Checks that each material claim matches the source text.
4. Delivers the answer with citations, or declines or asks a clarifying question if it fails.

See Understanding grounded answers for more on this loop.

IrisAgent also tests every model and prompt change against two eval sets: questions the AI should resolve, and questions it should decline. Changes that raise resolution at the cost of fabrication do not ship.

“Zero hallucinations” means unsupported claims are not delivered as grounded answers. IrisAgent prefers a decline, a clarifying question or a handoff over a confident guess. When automation rate and grounding conflict, grounding wins.

​

What you can configure

You make the engine more effective by improving its inputs and adding QA:

Area	What to do	Where
Knowledge coverage	Connect your KB, Confluence, public docs and uploads	Data sources, Upload Content
Product scope	Segment content by product so the wrong product’s articles are not used	Product Segmentation
Term alignment	Map customer wording to your KB vocabulary	Search Term Rewrites
Procedures	Enable or write procedures for common flows	Onboarding, Workflows
Post-delivery QA	Add AutoQA rules that flag made-up features, pricing or policies	AutoQA
Spot checks	Run Answer Grounding QA on a delivered answer	Support Analyst

A good first AutoQA rule (category Resolution Quality):

AI agent must not fabricate or hallucinate product features, pricing, or policies

​

Limits

• Stale or wrong articles still pass. An answer can be faithful to an outdated or off-topic article. Support Analyst labels this wrong article (a retrieval or KB problem).
• Grounded answers can be incomplete. An answer can cite the right article and still miss a step. Support Analyst labels this incomplete.
• KB gaps show up as declines. Fix them with new articles or Knowledge Suggestions (AutoKB), not by loosening the check.
• No always-on audit yet. Auditing every AI answer automatically is on the roadmap. Use AutoQA and Support Analyst spot checks today.
• Agents can edit suggestions. Copilot suggestions are starting points. Agents are accountable for what they send.

​

Related reading

• Understanding grounded answers
• Support Analyst: Answer Grounding QA
• AutoQA
• Connecting your helpdesk

Questions? [e-mail]