Inline Citation Insertion Repair Prompt

This prompt is designed for a post-processing step in a RAG or document-intelligence pipeline where a generated answer is semantically correct but its inline citation markers are missing, misplaced, or incorrectly anchored. The primary job-to-be-done is to repair the citation structure of an existing text without altering the answer's core meaning, claims, or prose style. The ideal user is an AI product engineer or platform developer who has already generated an answer from a model and now needs to ensure every factual assertion is properly and precisely linked to a provided set of source passages before the answer is shown to an end-user or saved to a system of record.

Use this prompt when you have a high-quality answer that fails a citation completeness or anchor-text validation check. This is a targeted repair tool, not a general-purpose answer generator. It requires two critical inputs: the original generated text and the exact source passages with their identifiers. The prompt works by instructing the model to map claims to sources and insert citation markers at the exact point of support, using anchor text from the source as a guide. It is most effective when the answer's claims are already grounded in the provided sources but the citation syntax is broken. Do not use this prompt if the answer contains hallucinated claims that cannot be found in any provided source; in that case, use a hallucination correction or answer regeneration prompt first.

Before deploying this prompt, ensure your pipeline has a clear definition of a valid citation format and a method for validating anchor text against source passages. The prompt includes a placeholder for a [CITATION_FORMAT] and [ANCHOR_VALIDATION_RULES], which you must replace with your system's specific requirements. This prompt is not a substitute for retrieval; it assumes the correct evidence is already in the context window. For high-stakes domains like healthcare or finance, always route the repaired output through a human review step and log the before/after citation state for auditability.

This prompt template is designed for post-processing pipelines that receive generated text with missing, misplaced, or malformed citation markers. It instructs the model to insert properly anchored inline citations by matching factual claims to provided source passages, without rewriting the answer's core meaning. The template uses square-bracket placeholders that you replace with your application's specific inputs, output schema, and constraints before sending to the model.

textCopy
You are a citation repair specialist. Your task is to take generated text that may have missing, misplaced, or malformed inline citation markers and insert properly anchored citations without altering the answer's semantic content.

## INPUT

### Generated Text to Repair
[GENERATED_TEXT]

### Source Passages with Citation Keys
[SOURCE_PASSAGES]

Each source passage includes a unique citation key in the format [SOURCE_ID] and the full passage text. Only cite sources provided here. Do not fabricate citations.

## INSTRUCTIONS

1. Read the generated text and identify every factual claim that requires source support.
2. For each claim, locate the best supporting source passage from the provided list.
3. Insert an inline citation marker immediately after the supported claim using the format [SOURCE_ID].
4. If a claim cannot be matched to any provided source passage, wrap it in [UNSUPPORTED]...[/UNSUPPORTED] tags.
5. Do not change the wording, structure, or meaning of the original text except to add citation markers and unsupported-claim tags.
6. Preserve all existing formatting, line breaks, and paragraph structure.

## OUTPUT FORMAT

Return a JSON object with this exact schema:

{
  "repaired_text": "string (the full text with citations inserted)",
  "citation_map": [
    {
      "citation_key": "string (the SOURCE_ID used)",
      "claim_snippet": "string (the exact text span the citation anchors to)",
      "source_passage_snippet": "string (the relevant excerpt from the source passage)"
    }
  ],
  "unsupported_claims": [
    {
      "claim_snippet": "string (the text wrapped in UNSUPPORTED tags)",
      "reason": "string (why no source passage matched)"
    }
  ],
  "repair_notes": ["string (any issues encountered during repair)"]
}

## CONSTRAINTS

- Never invent citation keys or source content.
- Never modify the original text's meaning, tone, or structure beyond citation insertion.
- If a claim is partially supported, cite the supporting portion and wrap only the unsupported portion in UNSUPPORTED tags.
- If the generated text contains existing citation markers that are correctly anchored, preserve them.
- If existing citation markers are misplaced, reposition them to the correct claim.
- Do not add citations to opinions, rhetorical questions, or transitional phrases.

## EXAMPLES

### Example 1: Missing Citation

Generated Text: "The company reported record revenue in Q3."
Source Passage: [SRC_001] "Acme Corp announced Q3 revenue of $4.2 billion, a company record."

Repaired: "The company reported record revenue in Q3[SRC_001]."

### Example 2: Unsupported Claim

Generated Text: "The CEO plans to acquire three competitors next year."
Source Passages: [SRC_002] "The CEO discussed expansion plans but did not name specific targets."

Repaired: "[UNSUPPORTED]The CEO plans to acquire three competitors next year.[/UNSUPPORTED]"

### Example 3: Misplaced Citation

Generated Text: "Patient outcomes improved[SRC_003] by 23% according to the study."
Source Passage: [SRC_003] "The study found a 23% improvement in patient outcomes."

Repaired: "Patient outcomes improved by 23% according to the study[SRC_003]."

## RISK LEVEL: [RISK_LEVEL]

[HIGH_RISK_INSTRUCTIONS]

Adapt this template by replacing the square-bracket placeholders with your application's specific values. The [GENERATED_TEXT] placeholder should receive the full text output from your upstream generation step. The [SOURCE_PASSAGES] placeholder expects a structured list of source objects, each containing a unique identifier and the full passage text. For high-risk domains such as healthcare, legal, or finance, set [RISK_LEVEL] to "high" and populate [HIGH_RISK_INSTRUCTIONS] with additional constraints like requiring human review of all [UNSUPPORTED] claims or mandating that citation anchors be verified against source passages before publishing. Always validate the output JSON against the schema before passing repaired text downstream, and log the citation_map and unsupported_claims fields for audit trails and citation coverage monitoring.

This prompt is designed to operate as a post-processing repair step within a broader RAG or document-intelligence pipeline. It should be invoked after an initial answer has been generated but before the response is returned to the user. The harness must provide the prompt with the original generated text, the complete set of source evidence chunks, and a strict output schema. Because this prompt modifies user-facing text, the harness must validate that the repair did not alter the semantic content of the answer or introduce hallucinated citations.

The implementation should follow a validate-repair-validate loop. First, run a lightweight structural check on the initial answer to detect missing or misplaced citation markers. If the check fails, invoke this repair prompt with the [ORIGINAL_TEXT] and [SOURCE_CHUNKS] placeholders populated. After receiving the repaired output, run a second validation pass that checks: (1) every inserted citation marker maps to a real source chunk ID, (2) no original claims were deleted or altered beyond citation insertion, and (3) the output matches the expected [OUTPUT_SCHEMA]. If validation fails, increment a retry counter and re-invoke the prompt with the validation errors included in the [CONSTRAINTS] field. Stop retrying after three attempts and escalate to a human review queue with the full trace.

For model choice, prefer a model with strong instruction-following and structured output capabilities, such as gpt-4o or claude-3-5-sonnet. Set temperature to 0 or a very low value (0.1) to minimize creative drift. The harness should log every invocation—including the original text, the source chunk IDs, the repaired output, and the validation results—as a structured trace. This trace is essential for debugging citation drift, auditing repair fidelity, and measuring the prompt's effectiveness over time. Use a trace ID that links the repair step back to the originating user request and retrieval call.

Avoid wiring this prompt directly into a synchronous user-facing response path without a timeout and circuit breaker. If the repair step takes longer than your p99 latency budget, return the original answer with a flag indicating that citation repair is pending, or fall back to a simpler post-hoc citation list appended to the response. Never silently drop the repair failure—always surface it in your observability dashboard so the team can tune retrieval quality, prompt instructions, or validation thresholds.