This prompt is for RAG system builders and citation-aware application engineers who need to move beyond topical relevance and identify exactly which retrieved passages contain directly quotable support for a claim. Its single job is to rank a provided retrieval set by citation worthiness: direct quote support, claim alignment, and source verifiability. Use this prompt when your application must produce answers with inline citations, when auditability of source grounding matters, or when you need to filter a large retrieval set down to the passages that actually contain claim-level evidence. The ideal user already has a retrieval pipeline producing candidate passages and needs a reliable ranking step before answer generation.

This prompt assumes you already have a retrieval set. It does not perform retrieval, rewrite queries, or generate the final answer. It does not handle multi-hop reasoning across passages or resolve contradictions between sources—those are separate concerns. The prompt is designed for single-step ranking: given a claim or question and a set of candidate passages, it returns a ranked list with explicit rationale for why each passage is or is not citation-worthy. This makes it suitable as a mid-pipeline component that feeds into a downstream answer generation step with strict citation requirements.

Do not use this prompt when you need raw topical relevance scoring without citation intent, when your retrieval set has not been pre-filtered for basic relevance, or when you need the model to generate the final answer directly. It is also not a replacement for retrieval itself—feeding it hundreds of passages without pre-filtering will produce noisy rankings and waste context budget. For high-stakes domains like legal or medical applications, always pair this prompt with human review of the ranked passages before they are used in a final answer. The ranking is a recommendation, not a verdict. If your application cannot tolerate any hallucinated citations, implement a downstream verification step that checks whether the cited passage actually contains the quoted claim.

This template is designed to be pasted directly into your prompt layer. It instructs the model to evaluate a set of retrieved evidence passages against a specific claim or question, ranking them by their suitability for citation in a final answer. The prompt prioritizes direct quote support, alignment with the claim, and the verifiability of the source. All dynamic inputs are represented as square-bracket placeholders, which your application must replace before sending the request to the model.

textCopy
You are an evidence-ranking specialist. Your task is to evaluate a set of retrieved evidence passages and rank them by their suitability for citation in a final answer to a user's question.

# INPUT

**User Question:** [USER_QUESTION]

**Claim to Support:** [CLAIM_TO_SUPPORT]

**Retrieved Evidence Passages:**
[EVIDENCE_PASSAGES]

# RANKING CRITERIA

Rank each passage on a scale of 1-5 (1=lowest, 5=highest) for each of the following criteria:

1.  **Direct Quote Support:** Does the passage contain a specific quote that directly supports the claim?
2.  **Claim Alignment:** How precisely does the evidence align with the core assertion of the claim?
3.  **Source Verifiability:** Is the source of the passage clearly identified and likely to be authoritative?

# CONSTRAINTS

- Do not rank a passage highly if it only tangentially relates to the claim.
- Penalize passages that are vague, lack specific details, or contain contradictory information.
- If a passage is entirely irrelevant, assign it a score of 0 for all criteria.
- Prioritize passages that provide direct, quotable evidence over those requiring inference.

# OUTPUT SCHEMA

Return a JSON object with a single key "ranked_evidence" containing an array of objects. Each object must have the following fields:

- `passage_id`: (string) The unique identifier of the passage from the input.
- `direct_quote_support_score`: (integer) 1-5
- `claim_alignment_score`: (integer) 1-5
- `source_verifiability_score`: (integer) 1-5
- `composite_score`: (number) The average of the three scores, rounded to one decimal place.
- `citation_worthiness`: (string) One of "high", "medium", "low", or "none".
- `rationale`: (string) A brief, specific explanation for the scores, referencing the passage content.

# EXAMPLE OUTPUT

```json
{
  "ranked_evidence": [
    {
      "passage_id": "doc_3_para_2",
      "direct_quote_support_score": 5,
      "claim_alignment_score": 5,
      "source_verifiability_score": 4,
      "composite_score": 4.7,
      "citation_worthiness": "high",
      "rationale": "Contains the direct quote '...the revenue increased by 20%...' which perfectly aligns with the claim. Source is a verified financial report."
    }
  ]
}

RISK LEVEL

[RISK_LEVEL]

To adapt this template, you must replace the placeholders. [USER_QUESTION] and [CLAIM_TO_SUPPORT] provide the necessary context for the model. [EVIDENCE_PASSAGES] should be a structured list of your retrieved documents, each with a unique ID and its text content. The [RISK_LEVEL] placeholder is a critical control; for high-stakes domains like healthcare or finance, you should append a final instruction such as: 'If the highest composite score is below 3.5, you must output an empty ranked_evidence array and state that no sufficiently strong evidence was found.' This prevents the model from fabricating a best-fit ranking from weak sources.

Before integrating this prompt into your application, test it with a golden dataset of claims and evidence where the correct ranking is known. Pay close attention to the citation_worthiness field in the output; this is your primary gating mechanism. In your application harness, you should parse the JSON output and only pass passages with a citation_worthiness of "high" or "medium" to the final answer generation step. Any other result should trigger a fallback, such as a request for more context or a refusal to answer.