Contradiction-Aware Answer Synthesis Prompt

This prompt is designed for the final answer generation step in a Retrieval-Augmented Generation (RAG) pipeline when the retrieved context contains conflicting information. Its core job is to prevent the model from fabricating a false consensus. Instead of smoothing over disagreements, the prompt instructs the model to synthesize an answer that explicitly surfaces contradictions, attributes each distinct position to its specific source passages, and refuses to present a unified narrative when the evidence is genuinely divided. The ideal user is a product engineer or AI architect building a research assistant, decision-support tool, or any user-facing application where transparency about evidential uncertainty is more valuable than a concise but misleading single answer.

You should use this prompt when your retrieval step returns multiple passages that disagree on a factual claim, methodology, or interpretation. It assumes you have already retrieved, ranked, and provided a set of passages to the model as [CONTEXT]. The prompt is particularly effective when the user's question is complex, the domain is nuanced (e.g., legal analysis, scientific research, market intelligence), and the end-user needs to understand the full landscape of evidence to make an informed decision. The output is not a simple string; it should be a structured object that separates the conflicting positions, their supporting citations, and a statement of the unresolved disagreement. This structure is critical for downstream UI rendering that can visually distinguish consensus from conflict.

Do not use this prompt when all retrieved passages agree, as the conflict-detection overhead will add unnecessary verbosity and may hallucinate minor stylistic differences as substantive disagreements. It is also the wrong tool when you need a single, definitive answer for a downstream API or automated workflow that cannot handle nuanced output. If the user expects a concise, direct response without source attribution, this prompt will produce an overly complex result that degrades the user experience. Finally, this prompt is not a substitute for a dedicated conflict detection step; it synthesizes an answer from known conflicts but is not optimized for exhaustively finding every contradiction in a large document corpus. For that, use a dedicated Evidence Discrepancy Flagging Prompt earlier in your pipeline.

This prompt template is designed for RAG application builders who need to generate answers from a set of retrieved passages that may contain conflicting information. Instead of forcing a single, falsely unified narrative, the prompt instructs the model to detect disagreements, attribute claims to specific sources, and present a balanced synthesis. The template uses square-bracket placeholders for your application's specific inputs, constraints, and output schema. Copy the block below into your system instructions or user message template, replacing each placeholder with your application data before sending the request to the model.

textCopy
You are a contradiction-aware answer synthesis engine. Your task is to produce a final answer to a user's question based on a set of provided source passages. These passages may agree, partially agree, or directly contradict each other.

## INPUT
- User Question: [USER_QUESTION]
- Retrieved Passages (with source IDs):
[RETRIEVED_PASSAGES]

## CONSTRAINTS
- Do not fabricate consensus. If sources disagree, you must acknowledge the disagreement explicitly.
- Attribute every factual claim to its source ID(s). Use inline citations: [Source: ID].
- If a source is silent on a claim, do not cite it for that claim.
- If the provided passages are insufficient to answer the question, state that clearly instead of guessing.
- Do not introduce external knowledge. Base your entire answer on the provided passages.
- [ADDITIONAL_CONSTRAINTS]

## OUTPUT FORMAT
Return a single JSON object with the following schema:
{
  "answer_summary": "A concise, plain-language summary of the best-supported answer, including a clear statement of the majority position if one exists.",
  "consensus_claims": [
    {
      "claim": "A factual statement supported by multiple sources.",
      "supporting_sources": ["source_id_1", "source_id_2"]
    }
  ],
  "contradictions": [
    {
      "topic": "The specific point of disagreement.",
      "position_a": {
        "claim": "The claim made by the first group of sources.",
        "sources": ["source_id_3"]
      },
      "position_b": {
        "claim": "The conflicting claim made by the second group of sources.",
        "sources": ["source_id_4"]
      },
      "resolution_notes": "A brief note on whether the conflict can be explained by scope, methodology, or recency differences, or if it remains unresolved."
    }
  ],
  "unsupported_aspects": ["Aspects of the question that the provided passages could not answer."],
  "confidence_level": "high|medium|low",
  "confidence_rationale": "A brief explanation for the confidence level, referencing source agreement, source quality, and evidence gaps."
}

## EXAMPLES
[FEW_SHOT_EXAMPLES]

To adapt this template, start by replacing [USER_QUESTION] with your end-user's query and [RETRIEVED_PASSAGES] with your retrieval pipeline's output, ensuring each passage has a unique, stable source ID. The [ADDITIONAL_CONSTRAINTS] placeholder allows you to inject domain-specific rules, such as preferring peer-reviewed sources or requiring a specific citation style. The [FEW_SHOT_EXAMPLES] placeholder is critical for complex schemas; provide 1-3 examples of the desired JSON output for questions with clear contradictions to calibrate the model's behavior. For high-stakes domains like legal or clinical review, add a constraint requiring the model to flag any contradiction for human review before presenting it to the user, and ensure your application layer validates the JSON schema before surfacing the output.

This prompt is designed to sit at the synthesis stage of a RAG pipeline, after retrieval and evidence ranking have produced a set of top passages. It expects pre-fetched context chunks with stable source identifiers, not raw search results. The harness should inject the ranked passages into the [CONTEXT] placeholder, each with a unique source ID, title, and snippet. The user question fills [INPUT], and [OUTPUT_SCHEMA] should specify a structured JSON format that includes an answer field, a contradictions array with claim, sources_for, sources_against, and resolution_note fields, and a citations array mapping each factual statement to its supporting source IDs. This structured output is critical for downstream validation and auditability.

Before the synthesized answer reaches a user, the harness must run two validation passes. First, a citation grounding check verifies that every source ID referenced in the output exists in the provided context and that no hallucinated sources appear. Second, a contradiction balance check ensures that when the prompt identifies a conflict, it represents both sides with at least one source each and does not silently collapse disagreement into false consensus. If either check fails, the harness should retry the prompt once with an explicit error message injected into [CONSTRAINTS]—for example, 'Your previous output referenced source S7 which was not provided. Use only the source IDs listed in the context.' After a second failure, escalate to a human review queue with the original context, the failed outputs, and the validation error details.

For model choice, use a model with strong instruction-following and structured output capabilities, such as GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro. Set temperature low (0.0–0.2) to reduce variance in contradiction detection. Enable structured output mode if the provider supports it, binding the response to the [OUTPUT_SCHEMA] JSON schema directly rather than relying on text parsing. Log every synthesis call with the prompt version, context passage IDs, model, and validation results. In high-stakes domains like legal or medical, always route outputs with detected contradictions to human review before publication, regardless of validation pass status. The harness should also track contradiction frequency over time—a sudden spike may indicate a retrieval quality issue or a source corpus update that introduced conflicting documents.