Knowledgebase Conflict Review with NLI

Objective

Two articles answer the same question differently — one says the Aurora plan includes email support, the other says it doesn't. Customers get whichever one search surfaces. Submit both passages as anchors for the claim under review and the conflict surfaces immediately.

Step 1: Verify against both articles at once

curl -s -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
  -d '{
    "model": "nli-default",
    "statement": "The Aurora plan includes email support.",
    "truth_anchors": [
      "The Aurora plan includes email support.",
      "The Aurora plan does not include email support."
    ],
    "closed_world": false,
    "contradiction_threshold": 0.6, "entailment_threshold": 0.6
  }' http://localhost:8080/v1/ai/nli/verify

Expected (verified on v2.1.1-ce): verdict LIE naming the denying article as the violated anchor (entailment 1.00 AND contradiction 1.00 present) — a qualifying contradiction wins even when another anchor fully supports the claim. That asymmetry is the point: a conflict is a content defect either way, and the safe behavior is to flag it.

Step 2: Order must not change the verdict

curl -s -H "Authorization: Bearer $TOKEN" -H 'Content-Type: application/json' \
  -d '{
    "model": "nli-default",
    "statement": "The Aurora plan includes email support.",
    "truth_anchors": [
      "The Aurora plan does not include email support.",
      "The Aurora plan includes email support."
    ],
    "closed_world": false,
    "contradiction_threshold": 0.6, "entailment_threshold": 0.6
  }' http://localhost:8080/v1/ai/nli/verify

Expected (verified): the same verdict and the same violated anchor; all_scores follows the new input order. Anchor order is presentation, not priority — except on exact score ties, where the earliest supplied anchor is selected.

Use it from your app

Run it as a nightly content-QA sweep: for each claim in a canonical FAQ list, gather the passages that mention it and verify. A flagged conflict goes to an editor with both article IDs — the classifier does not decide which article is authoritative, handle effective dates, or retire the obsolete one. Keep article provenance outside the request.

Key Concepts Learned

• Contradiction-priority policy makes conflict detection conservative by design.
• NLI finds that articles disagree; which one is right is editorial judgment.