Citation Completeness Check Prompt

This prompt is designed for compliance reviewers, audit preparation teams, and AI pipeline engineers who need to verify that every factual claim in a generated text is backed by a citation. It checks for orphan claims, assesses whether the provided citation set is sufficient for the claims made, and produces a structured completeness report with a gap analysis. Use this before submitting AI-assisted outputs for regulatory review, board reporting, or any workflow where an unsupported factual statement creates audit risk.

This prompt assumes you already have a text with claims and a set of citations. It does not verify whether the citations are accurate or whether the sources actually support the claims; pair it with the Citation Verification Prompt Template for that step. The ideal input is a finalized or near-final draft where the author has already attempted to cite their sources. The prompt works best when the input text contains discrete, verifiable factual claims rather than vague commentary or opinion. It is not designed for real-time streaming outputs, conversational chat, or creative writing where factual grounding is not expected.

Do not use this prompt as a substitute for human review in high-stakes regulated filings. It identifies gaps in citation coverage but cannot assess whether the missing citations would have contradicted the claims. For workflows involving financial disclosures, clinical documentation, or legal submissions, treat the completeness report as a triage tool that directs human reviewers to the highest-risk gaps. Also avoid using this prompt on texts where citations are intentionally sparse, such as executive summaries or marketing narratives, unless you have first defined a clear citation policy for that document type.

This template is designed to be pasted directly into your orchestration layer. It accepts a text containing factual claims, a set of citations or source documents, and optional constraints such as a domain-specific citation policy or a minimum support threshold. The prompt instructs the model to produce a structured completeness report identifying uncited claims, under-supported claims, and gaps where additional evidence is required. Replace every square-bracket placeholder with your actual data before sending the request.

textCopy
You are a citation completeness auditor. Your job is to check whether every factual claim in the provided text has been cited and whether the provided citations are sufficient to support the claims made.

## INPUT
[INPUT]

## CITATIONS AND SOURCE MATERIAL
[CITATIONS]

## OUTPUT SCHEMA
Return a JSON object with the following structure:
{
  "overall_completeness": "COMPLETE" | "PARTIAL" | "INSUFFICIENT",
  "summary": "One-paragraph summary of citation coverage and the most important gaps.",
  "claims": [
    {
      "claim_text": "The exact factual claim from the input.",
      "citation_status": "CITED" | "UNCITED" | "UNDER_SUPPORTED",
      "supporting_citations": ["List of citation IDs or source references that support this claim."],
      "gap_description": "If UNCITED or UNDER_SUPPORTED, explain what is missing and what kind of evidence would be needed.",
      "severity": "HIGH" | "MEDIUM" | "LOW"
    }
  ],
  "gap_analysis": {
    "uncited_claims_count": 0,
    "under_supported_claims_count": 0,
    "critical_gaps": ["List of the most important gaps that require immediate attention."],
    "recommended_actions": ["Specific actions to resolve gaps, such as retrieving additional sources or flagging for human review."]
  }
}

## CONSTRAINTS
[CONSTRAINTS]

## RISK LEVEL
[RISK_LEVEL]

## INSTRUCTIONS
1. Extract every factual claim from the input text. A factual claim is any statement that asserts something about the world, including statistics, dates, events, attributions, comparisons, and descriptions of entities.
2. For each claim, determine whether it is supported by at least one provided citation or source document.
3. If a claim is cited, assess whether the cited source actually supports the claim. Flag claims where the citation is present but the source does not support the claim as UNDER_SUPPORTED.
4. If a claim has no citation, flag it as UNCITED.
5. Assign a severity level to each gap: HIGH for claims that are central to the text and unsupported, MEDIUM for supporting claims with weak evidence, LOW for minor or tangential claims.
6. If [RISK_LEVEL] is HIGH, include a note that all UNCITED and UNDER_SUPPORTED claims must be reviewed by a human before publication.
7. Do not invent citations. If a claim is uncited, report it as uncited.
8. If [CONSTRAINTS] includes a domain-specific citation policy, apply those rules when assessing sufficiency.

Adapt this template by adjusting the output schema to match your downstream consumption format, adding domain-specific citation policies in the [CONSTRAINTS] block, and setting [RISK_LEVEL] to control whether the output includes mandatory human-review flags. For high-stakes domains such as healthcare, legal, or financial compliance, set [RISK_LEVEL] to HIGH and ensure that the gap analysis feeds into a human review queue before any output is published. Wire the JSON output into a validation step that checks for structural completeness before passing results to your review or escalation system.

The Citation Completeness Check Prompt is not a one-off chat interaction; it is a verification stage in a production pipeline. The prompt expects a structured input containing the full generated text and a list of citations with their source content. The output is a completeness report with gap analysis. To integrate this reliably, you must wrap the prompt in a harness that validates inputs, parses the structured output, and routes the results to the appropriate downstream system—whether that is a review queue, a retry loop, or an audit log.

Start by defining a strict input schema. The prompt requires [FULL_TEXT] (the generated answer or document), [CITATIONS] (an array of objects, each with citation_id, source_text, and optional source_metadata), and [CLAIM_TYPES] (a list of claim categories that require citation, such as statistical, legal, temporal, or attributed_opinion). Validate that every citation object includes retrievable source text before calling the model; an empty source field will produce a false positive for missing evidence. On the output side, parse the model response against a JSON schema that includes claim_list (each claim with claim_id, claim_text, claim_type), citation_gaps (claims with no supporting citation), sufficiency_assessment (per-citation coverage score), and overall_completeness_score. Use a library like jsonschema or pydantic to validate the parsed output. If parsing fails, retry with a stricter [OUTPUT_SCHEMA] instruction or fall back to a repair prompt from the Output Repair and Validation Prompts pillar.

For compliance workflows, route the completeness report through a human review step when the overall_completeness_score falls below a configurable threshold (e.g., 0.85) or when any citation_gap is flagged with severity: high. Log every verification run with the input hash, model version, timestamp, and completeness score. This creates an audit trail that downstream governance tools can consume. If you are using a RAG pipeline, feed the gap analysis back into the retrieval stage: claims flagged as under-supported should trigger additional retrieval with the claim text as a query. Avoid using this prompt on streaming outputs; the completeness check requires the full text and full citation set to produce a meaningful gap analysis. Model choice matters: use a model with strong instruction-following and JSON mode support, such as gpt-4o or claude-3-5-sonnet, and set temperature=0 to minimize variance in the gap detection logic.

The most common production failure is a mismatch between the claims the model extracts and the claims a human reviewer would identify. Mitigate this by running the prompt against a golden dataset of 20–50 annotated examples where claim boundaries and citation gaps are labeled by domain experts. Measure precision and recall on gap detection, not just overall completeness score agreement. If recall drops below 0.90, add few-shot examples of missed gap patterns to the [EXAMPLES] block. Do not deploy this prompt as a blocking gate until you have calibrated the completeness threshold against your domain's risk tolerance. Start in shadow mode, logging completeness reports alongside existing review processes, and compare the prompt's gap flags against human-identified gaps for at least two review cycles before cutting over.