Citation Verification

Verify that your agent's claims are actually grounded in its sources.

Overview

When an agent generates a response using RAG or deep research, citation verification checks each factual claim against the source documents. Claims are scored using cosine similarity between embeddings, producing per-claim verdicts: supported, weak, unverifiable, or contradicted.

Quick Start

Enable Automatic Verification

Add to your agent.config.json:

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

Deep research responses (depth: "deep") are always verified. With verifyCitations: true, moderate-depth queries also verify.

Use the CLI

# Verify citations during deep research
wunderland chat --provider openai

# Ask a factual question — citations verified automatically
> What is the population of Tokyo?

Help Topic

wunderland help citations

How It Works

1. Claim extraction — the response is split into atomic factual claims
2. Batch embedding — all claims and source documents are embedded in one call
3. Cosine similarity — each claim is compared against each source
4. Verdict assignment — based on similarity thresholds:
   • supported (>= 0.6) — claim matches a source
   • weak (0.3 - 0.6) — partial match
   • unverifiable (< 0.3) — no source matches
   • contradicted — NLI detects contradiction

On-Demand Verification

Agents can explicitly verify text using the verify_citations tool:

You: "Check if this summary is accurate"
Agent calls: verify_citations({
  text: "The Earth is 4.5 billion years old. It orbits the Sun every 365.25 days.",
  sources: [{ content: "Earth formed approximately 4.54 billion years ago." }],
  webFallback: true
})

When webFallback: true, claims that can't be verified against provided sources are checked via web search (requires SERPER_API_KEY or TAVILY_API_KEY).

The fact-grounding Skill

Install the fact-grounding skill to make your agent automatically verify claims:

wunderland init my-agent --preset research-assistant

Or add to your agent config:

{
  "skills": ["fact-grounding"]
}

The skill instructs the agent to:

• Verify key factual claims before presenting
• Mark unverified claims with "[unverified]"
• Cite sources inline: "According to [Source]..."
• Flag contradictions with both sides

Integration with Deep Research

Citation verification plugs into the deep research pipeline:

Query → Decompose → Search (7 providers) → Reranker Chain → Synthesis → Citation Verification

The reranker chain ensures the best sources reach synthesis. Citation verification then confirms the synthesized claims match those sources.

Configuration Reference

Config Key	Default	Description
queryRouter.verifyCitations	false	Enable automatic verification
supportThreshold	0.6	Cosine similarity for "supported"
unverifiableThreshold	0.3	Below this = "unverifiable"

Environment Variables

Variable	Purpose
SERPER_API_KEY	Web fallback for unverifiable claims
TAVILY_API_KEY	Alternative search for web fallback
COHERE_API_KEY	Reranking before citation (optional)

Related

• wunderland help citations — CLI quick reference
• wunderland help retrieval — RAG and search pipeline
• wunderland help security — guardrails including grounding guard