Agent Output Conflict Resolution for RAG Citation Conflicts Prompt Template

This prompt is designed for RAG system builders and evaluation engineers who operate multi-agent pipelines where two or more agents produce answers supported by conflicting citations. When Agent A cites Source X to claim a fact and Agent B cites Source Y to contradict it, the system must resolve the conflict rather than presenting users with contradictory information or silently picking a winner. The prompt acts as an arbitration layer that receives the conflicting agent outputs, their cited sources, and any available source metadata, then produces a citation-resolved answer with explicit source authority weighting, contradiction acknowledgment, and evidence quality notes.

Use this prompt when your RAG pipeline includes multiple retrieval or generation agents that may surface incompatible evidence, when you need an auditable resolution trail, and when source trustworthiness varies by recency, authority, or domain relevance. The prompt expects structured inputs: the conflicting agent outputs, the specific citations each agent relied on, and source metadata such as publication date, author credentials, domain authority score, and retrieval rank. It returns a resolved answer that preserves areas of agreement, explicitly flags contested claims, attributes each resolved position to its supporting evidence, and provides a confidence score for the resolution. Wire this prompt after agent outputs are collected and normalized, but before the final answer reaches the user. Validate the output by checking that every resolved claim has a citation, that contradictions are acknowledged rather than erased, and that the resolution reasoning is traceable back to the input evidence.

Do not use this prompt for single-agent RAG outputs with no conflict, for conflicts that are purely stylistic rather than factual, or when the underlying sources are unavailable for verification. If the sources themselves are inaccessible or the agents cannot produce the specific text they relied on, this prompt will generate plausible but unverifiable resolutions—a dangerous failure mode. For high-stakes domains such as healthcare, legal, or financial applications, always route the prompt's output through human review before it reaches end users. The prompt is not a substitute for source verification; it is an arbitration layer that makes conflicts explicit and resolution reasoning auditable. If the conflict involves more than three agents or more than ten contested claims, consider splitting the input into smaller arbitration units to avoid resolution quality degradation.

This template resolves conflicts when multiple RAG agents return different answers backed by different citations. It forces the model to compare sources, assess authority and recency, and produce a single answer that explicitly acknowledges contradictions rather than silently discarding minority evidence. Copy the template below and replace every square-bracket placeholder with your actual data before sending it to the model.

textCopy
You are a citation conflict arbitrator for a RAG system. Your job is to resolve disagreements between agent outputs that cite different sources, and produce a single evidence-grounded answer.

## INPUTS

### User Question
[USER_QUESTION]

### Agent Outputs with Citations
[AGENT_OUTPUTS]

Each agent output includes:
- The agent's answer or claim
- A list of citations with source identifiers, publication dates, and authority scores
- The agent's self-reported confidence

### Source Authority Rules
[SOURCE_AUTHORITY_RULES]

### Output Schema
You must return valid JSON matching this schema:
{
  "resolved_answer": "string (the final answer, citing specific sources inline)",
  "contradictions_acknowledged": [
    {
      "claim": "string (the contested claim)",
      "position_a": "string (first agent's position with source)",
      "position_b": "string (second agent's position with source)",
      "resolution": "string (how you resolved it and why)"
    }
  ],
  "evidence_quality_notes": [
    {
      "source_id": "string",
      "quality_issue": "string (recency, authority, relevance, or consistency concern)",
      "impact": "string (how this affected the final answer)"
    }
  ],
  "unresolved_issues": ["string (any contradictions that could not be resolved)"],
  "confidence": "high|medium|low",
  "escalation_recommended": true|false
}

## CONSTRAINTS
[CONSTRAINTS]

## RESOLUTION RULES
1. Prefer sources with higher authority scores unless a lower-authority source is significantly more recent and directly addresses the question.
2. When sources directly contradict, acknowledge the contradiction explicitly. Do not silently discard minority evidence.
3. If no source is clearly superior, report the disagreement as unresolved and recommend escalation.
4. Flag any source that appears fabricated, miscited, or misrepresented.
5. If all agents agree but cite different sources, note the source diversity as a strength.
6. Never invent citations. Only reference sources provided in the agent outputs.

## EXAMPLES
[EXAMPLES]

Adaptation notes: Replace [AGENT_OUTPUTS] with the structured outputs from your RAG agents, including their answers, citations, and confidence scores. [SOURCE_AUTHORITY_RULES] should define how your system weights sources (e.g., internal docs > public web, peer-reviewed > blog posts, recency thresholds). [CONSTRAINTS] can include domain-specific rules like regulatory citation requirements or forbidden source types. [EXAMPLES] should contain at least one example where sources conflict and one where they agree but cite differently. If your use case involves regulated content, add a constraint requiring human review when escalation_recommended is true.

What to watch for: The model may still fabricate citations if agent outputs are ambiguous about source identifiers. Always validate that every cited source in the resolved answer appears in the original agent outputs. If you are operating in a high-risk domain, route outputs with confidence: low or escalation_recommended: true to a human reviewer before surfacing the answer to users. Run periodic eval checks comparing resolved answers against ground-truth evidence to catch hallucinated resolutions.

This prompt is designed to sit at the end of a multi-agent RAG pipeline where two or more agents have produced answers citing different or contradictory sources. The harness must receive the original user query, the conflicting agent outputs (each with their citations), and the full text of all cited passages. Before calling the LLM, validate that every citation in the agent outputs resolves to a passage in the provided evidence set—if a citation points to a missing passage, either fetch it or flag the agent output as containing an unresolvable citation before arbitration begins.

Wire the prompt into an application layer that performs pre-call validation, post-call verification, and configurable retry logic. On the input side, check that [AGENT_OUTPUTS] contains at least two entries with distinct citations, that [EVIDENCE_PASSAGES] includes all cited sources, and that [SOURCE_AUTHORITY_MAP] assigns a numeric weight to each source. On the output side, parse the JSON response and run three verification checks: (1) every citation in the resolved answer maps to a passage in the evidence set, (2) the contradictions_acknowledged array includes every pair of sources the prompt identified as conflicting, and (3) the evidence_quality_notes field is non-empty when any source authority score falls below the configured threshold. If any check fails, increment a retry counter and resubmit with the validation error appended to [CONSTRAINTS]. After three failed retries, route the entire package—original query, agent outputs, evidence, and failed resolution attempts—to a human review queue with a structured handoff that highlights which validation rule kept failing.

For production observability, log the full input payload, the resolved output, the verification results, and the retry count on every invocation. Attach a trace ID that links this arbitration step to the upstream agent calls and the downstream consumer of the resolved answer. When the escalation_required field is true or the retry limit is hit, the human review hook must present the reviewer with a side-by-side view of the conflicting agent outputs, the evidence passages, and the partial resolution the model produced before failure. Avoid deploying this prompt without citation-verification guards—RAG pipelines that skip post-hoc citation validation will silently propagate fabricated references into the final answer.