Citation Verifier

Verify claims in agent responses against sources using semantic similarity with optional web fallback.

Overview

The citation verifier decomposes text into atomic claims, embeds them alongside source documents, and uses cosine similarity to determine whether each claim is supported by a source. Claims that can't be verified against provided sources can optionally be checked via web search.

Package: @framers/agentos-ext-citation-verifier

When Citations Are Verified

Research Depth	Automatic Verification	Config Override
none / quick	No	verifyCitations: true in config
moderate	No	verifyCitations: true in config
deep	Yes (automatic)	Always on
On-demand	Agent calls verify_citations tool	N/A

Tool: verify_citations

verify_citations({
  text: "The Earth orbits the Sun at ~150 million km.",
  sources: [{ content: "Earth's orbital distance averages 149.6 million km." }],
  webFallback: true
})

Parameters

Parameter	Type	Required	Description
text	string	Yes	Text containing claims to verify
sources	array	No	Sources to check against
webFallback	boolean	No	Search web for unverifiable claims

Output

{
  "claims": [
    {
      "text": "The Earth orbits the Sun at ~150 million km.",
      "verdict": "supported",
      "confidence": 0.87,
      "source": { "snippet": "Earth's orbital distance averages 149.6 million km." }
    }
  ],
  "totalClaims": 1,
  "supportedCount": 1,
  "supportedRatio": 1.0,
  "summary": "1/1 claims verified (100%)"
}

Verdicts

• supported (similarity >= 0.6) — claim matches a source
• weak (similarity 0.3-0.6) — partial match, lower confidence
• unverifiable (similarity < 0.3) — no source matches
• contradicted — NLI model detects contradiction with source

Configuration

agent.config.json

{
  "queryRouter": {
    "verifyCitations": true
  }
}

When enabled, responses from moderate and deep research queries include a grounding field with per-claim verdicts.

Core API (Programmatic)

import { CitationVerifier } from '@framers/agentos';

const verifier = new CitationVerifier({
  embedFn: async (texts) => embeddingManager.embed(texts),
  supportThreshold: 0.6,
  unverifiableThreshold: 0.3,
});

const result = await verifier.verify(
  "Tokyo has a population of 14 million.",
  [{ content: "Greater Tokyo metropolitan area: 37 million. Tokyo proper: 14 million." }]
);
// result.claims[0].verdict === 'supported'

Related

• fact-grounding skill — instructs agents to verify claims before presenting
• grounding-guard guardrail — real-time NLI streaming verification
• fact_check tool — web-based single-claim verification
• Reranker chain — multi-stage result ranking before citation