Cross-Document Citation Repair Prompt Template

This prompt is a post-generation repair tool for multi-document RAG systems where citations point to the correct chunk of text but the wrong parent document. It is designed for the compliance audit and remediation loop, not for generating initial answers. The ideal user is an engineering lead or MLOps engineer who has already observed document-level attribution errors in production logs or eval results and needs a structured, automatable correction step before the answer is surfaced to users or written to an audit log.

Use this prompt when your retrieval corpus contains documents with overlapping or semantically similar content, causing the model to confuse which document a chunk belongs to. It is also appropriate when chunk-level citation IDs are ambiguous across documents, or when a downstream compliance requirement mandates document-level attribution accuracy. The prompt requires three inputs: the original answer with inline citations, the cited source chunks with their chunk IDs, and the full document corpus metadata including document IDs, titles, and any other distinguishing fields. It outputs a structured correction log that remaps each citation to the correct document and flags any citations that could not be resolved.

Do not use this prompt to fix chunk-level misalignment within a single document, to generate initial answers from retrieved context, or to repair hallucinated citations where the cited chunk itself does not exist. It assumes the chunk-level evidence is correct and only the document assignment is wrong. For those other failure modes, use the RAG Citation Answer Repair, Citation-to-Evidence Alignment, or Hallucinated Citation Removal prompts instead. Before deploying this prompt in production, pair it with a validation step that checks the correction log against the corpus metadata to prevent the repair loop from introducing new errors.

This prompt template is designed to be the core repair step in a post-generation validation pipeline. It receives a model's original answer, the set of source documents it was supposed to use, and a list of detected citation errors. Its job is to produce a corrected answer with citations remapped to the correct source documents, along with a structured correction log. The template uses square-bracket placeholders for all dynamic inputs—replace each one with actual values from your application context before sending the prompt to the model.

textCopy
You are a citation repair specialist. Your task is to correct cross-document citation errors in a model-generated answer.

## INPUT

### Original Answer with Broken Citations
[ORIGINAL_ANSWER]

### Source Document Corpus
[DOCUMENT_CORPUS]

### Detected Citation Errors
[ERROR_LIST]

## TASK

1. For each citation error in the error list, identify the claim that was incorrectly attributed.
2. Search the source document corpus to find which document actually contains the supporting evidence for that claim.
3. Remap the citation to the correct document ID.
4. If no document in the corpus supports the claim, replace the citation with [UNSUPPORTED] and flag it in the correction log.
5. Preserve all other text, formatting, and correct citations exactly as they appear in the original answer.

## OUTPUT_SCHEMA

Return a JSON object with this structure:
{
  "repaired_answer": "The full answer text with corrected citations",
  "correction_log": [
    {
      "original_citation": "string",
      "original_document_id": "string",
      "corrected_document_id": "string or null if unsupported",
      "claim_text": "string",
      "correction_type": "remap|removal|unresolvable",
      "confidence": "high|medium|low"
    }
  ],
  "summary": {
    "total_citations": "number",
    "errors_corrected": "number",
    "unsupported_claims_flagged": "number",
    "unresolvable_errors": "number"
  }
}

## CONSTRAINTS

- Do not alter the original answer text except to change citation markers.
- Only use document IDs that exist in the provided source document corpus.
- If a claim appears in multiple documents, cite the most directly relevant one.
- Set confidence to "low" if the supporting evidence is partial or ambiguous.
- Never invent document IDs, source content, or supporting evidence.

Adaptation guidance: Replace [ORIGINAL_ANSWER] with the full text output from your RAG system, including all inline citation markers. [DOCUMENT_CORPUS] should contain every source document with its unique ID and full text—structure this as a JSON array or formatted list that the model can search against. [ERROR_LIST] should be a structured list of detected mismatches, typically produced by an upstream validation step that compares cited document IDs against the actual corpus. For high-stakes domains, add a [RISK_LEVEL] placeholder that gates whether corrections are applied automatically or routed to human review. The output schema is designed to be machine-readable so your application can parse the correction log and apply only high-confidence remappings automatically while escalating low-confidence or unresolvable cases.

The Cross-Document Citation Repair prompt operates as a post-retrieval correction layer, sitting between the initial answer generation step and the final user-facing output. In a typical RAG pipeline, a model generates an answer with inline citations referencing source chunks, but when the retrieval corpus spans multiple documents, the model can misattribute a citation to the wrong document—citing doc_7 when the evidence actually came from doc_3. This prompt accepts the generated answer, the full set of retrieved source chunks with their document-level metadata, and a mapping of citation markers to their claimed source IDs. It then produces a corrected answer with remapped citations and a structured correction log that records every attribution change, the evidence that justified the change, and a confidence score per correction.

To integrate this into an application, wrap the prompt in a repair loop that triggers after the primary answer generation step. The harness should: (1) extract all citation markers from the generated answer using a regex or parser that matches your citation format; (2) retrieve the full text of each cited source chunk from your vector store or document store, including the parent document identifier; (3) call the repair prompt with the answer, the source chunk map, and a [CITATION_FORMAT] specification; (4) parse the structured output—typically a JSON object containing repaired_answer, correction_log, and unresolvable_citations; (5) validate that every citation in the repaired answer maps to a real source chunk in the provided corpus; and (6) if unresolvable_citations is non-empty, either escalate for human review or apply a fallback strategy such as removing the unresolvable citation and inserting an abstention marker. For high-throughput systems, batch multiple answers into a single repair call where context windows permit, but keep each answer's source chunk set isolated to avoid cross-contamination.

Model choice matters here. This task requires strong instruction-following and structured output reliability, not creative generation. Use a model with native JSON mode or structured output support—GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro with response schema constraints. Set temperature=0 to minimize variance in citation remapping. Implement a retry policy with up to 3 attempts if the output fails JSON schema validation or if the correction log references source IDs not present in the input corpus. Log every repair call with: the original answer, the correction log, the validation result, and whether the output was served to the user or escalated. This audit trail is essential for debugging attribution drift over time and for demonstrating compliance in regulated domains. Avoid using this prompt on streaming outputs without first buffering the complete answer—partial citation repair on incomplete text will produce incorrect remappings and false-positive unresolvable flags.