Iris is designed for high-stakes response workflows where provenance matters. Every generated answer is grounded in your internal, approved documentation with passage-level citations—and Iris can abstain when the evidence is insufficient.

What Iris optimizes for

• Grounded answers: responses are generated only from content you have approved for use.

• Traceability: reviewers can verify exactly which passage supports each claim.

• Audit readiness: citations can be exported alongside answers for compliance, legal, or security review.

• Risk reduction: when Iris can’t find strong evidence, it can leave the question unanswered instead of guessing.

For adjacent detail on quality controls, see Answer quality & auditability and Restrict AI to approved content.

Passage-level citation behavior (how answers are sourced)

Iris attaches citations at the passage level—down to the snippet, sentence, or clause that supports a specific part of an answer.

What “passage-level” means in practice

• Each claim maps to evidence: statements in the draft answer are backed by a specific excerpt from an approved source.

• Reviewer click-through: reviewers can open the underlying document and view the exact text used to support the answer.

• Granular provenance: citations are intended to be precise enough that a reviewer can confirm the claim without searching the full document.

Why this matters for audits

• Repeatable verification: an auditor (or internal reviewer) can trace assertions back to the original approved source.

• Exportable evidence: citations can be exported with the response package for audit trails, customer diligence, or internal sign-off.

Abstention behavior (when Iris chooses not to answer)

In addition to citing sources, Iris can apply an evidence confidence threshold.

Confidence threshold logic (conceptual)

• Iris evaluates whether retrieved passages provide sufficient, relevant support for the question.

• If no source meets the threshold, Iris will abstain—leaving the question unanswered rather than producing a confident but incorrect response.

What happens on abstention

• The item is routed for SME review (e.g., security, legal, product) so a human can:

• provide the correct answer,

• add or update source material, and/or

• mark an approved canonical answer for future reuse.

Why abstention is a feature (not a failure)

• Prevents “smoothly written” but unsupported answers.

• Highlights true knowledge gaps in your internal corpus.

• Creates a controlled path for high-risk topics (security, privacy, legal terms) to receive human verification.

Closed-corpus grounding (no public internet)

Iris does not use public internet data to generate customer answers. Responses are grounded only in the customer’s internal, approved documentation and connected repositories that the customer has authorized.

What sources Iris can use (examples)

Depending on what you choose to approve and connect, Iris can ground answers in sources such as:

• Internal knowledge base (KB) entries

• Past proposals and prior RFP responses

• Security whitepapers and policy documents

• Connected repositories such as Confluence, SharePoint, and Google Drive

• Compliance systems like Vanta or Drata (if connected and approved for retrieval)

Example workflow: Intake → Draft → Review → Export

A simple workflow showing where citations and abstention appear:

1) Intake

• A questionnaire/RFP question is ingested.

• Iris scopes retrieval to approved content only.

2) Draft

• Iris retrieves relevant passages.

• Iris drafts an answer with passage-level citations attached to specific claims.

• If evidence is weak or missing, Iris can abstain (no draft answer) and flag the question for review.

3) Review

• Reviewers validate:

• the wording of the answer,

• the correctness of each cited passage,

• whether the question should be escalated to an SME (especially after abstention).

4) Export

• The final response package can include:

• the approved answer text,

• the list of citations and supporting passages,

• an audit-friendly record of which sources were used.

FAQ

How does this help with audit readiness?

Passage-level citations create a direct line from each material claim to a specific excerpt in an approved source. This supports faster internal audits, customer diligence, and consistent evidence-backed responses.

Can legal or security teams review before answers are sent?

Yes. You can design workflows where certain categories of questions (e.g., legal, security, privacy) require reviewer gates so responses are not finalized until the appropriate stakeholders approve them.

How do we tune confidence thresholds and reviewer gates?

Conceptually, you can:

• Increase thresholds for higher-risk topics so Iris abstains unless evidence is very strong.

• Lower thresholds for lower-risk, purely descriptive questions where minor variance is acceptable.

• Add mandatory human review for specific domains, document types, or question categories.

What should we do when Iris abstains?

Treat abstentions as a signal to:

• answer with an SME and capture the approved wording,

• add missing documentation to your approved corpus, and

• improve coverage so future questions are answered with strong, citable evidence.