Citation Drift Detection and Correction Prompt

Use this prompt when your system generates multi-paragraph answers with inline citations and you need to verify that each citation marker still accurately reflects the source passage it claims to reference. Citation drift is a common failure mode in RAG pipelines where a model correctly cites a source for an initial claim but then continues to use that same citation marker for subsequent claims that are not supported by the original passage. This prompt detects that drift, measures the semantic distance between the cited passage and the claim, and re-anchors citations to the correct sources. It is designed for AI product teams building cited answer systems where unattributed or misattributed claims create compliance risk and erode user trust.

This prompt assumes you already have a generated answer with inline citation markers and the full text of each cited source passage available for comparison. The ideal workflow runs this prompt as a post-generation validation step before the answer reaches the user. You should not use this prompt when the generated answer contains no citations, when source passages are unavailable, or when the answer is a single sentence where drift is unlikely. For those cases, use the Missing Citation Retry Prompt or the Unsupported Claim Correction Prompt instead. This prompt is also not a substitute for retrieval quality improvements—if your retriever consistently returns irrelevant passages, fix retrieval before applying drift correction.

The prompt works by extracting each claim-citation pair from the generated text, computing a drift distance score that measures how well the cited passage supports the claim, and then either re-anchoring the citation to a more appropriate source from your evidence set or flagging the claim as unsupported. You control the drift threshold that triggers correction. Set it conservatively (high semantic similarity required) for compliance-sensitive domains like legal or medical text, and more permissively for exploratory research summaries. Always log drift scores and correction actions for audit trails, and consider routing claims with borderline scores to human review rather than automatic re-anchoring.

Before deploying this prompt in production, establish a golden dataset of answers with known citation drift patterns and measure correction accuracy. Common failure modes include over-correction (re-anchoring to a source that is topically related but factually different), under-correction (missing subtle drift where the model shifts from 'may cause' to 'causes'), and cascading errors where correcting one citation breaks the logical flow of surrounding claims. Wire the prompt into a validation harness that compares pre- and post-correction answers, logs every change, and triggers a human review queue when the number of corrections exceeds a configurable threshold or when drift scores fall into an ambiguous range.

This template is designed to be pasted directly into your citation verification harness. It instructs the model to compare each citation in a long-form generated answer against its intended source passage, measure the semantic distance of any drift, and re-anchor the citation to the most faithful supporting passage. The prompt uses strict square-bracket placeholders that you must replace with your own data before execution.

textCopy
You are a citation drift detection and correction system. Your task is to analyze a long-form generated answer and its associated source passages to identify where citations have drifted away from their intended evidence.

## INPUT
- Generated Answer: [GENERATED_ANSWER]
- Source Passages (with IDs): [SOURCE_PASSAGES]
- Citation Map (citation marker to intended source ID): [CITATION_MAP]

## OUTPUT_SCHEMA
Return a JSON object with the following structure:
{
  "drift_analysis": [
    {
      "citation_marker": "string",
      "intended_source_id": "string",
      "claimed_claim": "string",
      "actual_source_content": "string",
      "drift_detected": boolean,
      "drift_distance": "none" | "minor" | "moderate" | "severe",
      "drift_description": "string",
      "corrected_source_id": "string or null",
      "corrected_source_passage": "string or null",
      "requires_regeneration": boolean
    }
  ],
  "overall_drift_score": number,
  "recommendations": ["string"]
}

## CONSTRAINTS
- Compare the claim associated with each citation marker against the content of its intended source passage.
- If the claim is not supported by the intended passage, search all provided source passages for the best supporting evidence.
- Classify drift distance as "none" if the claim is directly supported, "minor" if the claim is a reasonable paraphrase, "moderate" if the claim adds unsupported details, and "severe" if the claim contradicts or invents content not in any source.
- If no source passage supports the claim, set corrected_source_id to null and requires_regeneration to true.
- Do not invent sources. Only use source IDs present in the provided passages.
- overall_drift_score should be a float between 0.0 (no drift) and 1.0 (complete drift across all citations).

## EXAMPLES
[EXAMPLES]

## RISK_LEVEL
[RISK_LEVEL]

To adapt this template, replace [GENERATED_ANSWER] with the full text of the model's output, including inline citation markers. [SOURCE_PASSAGES] should contain the retrieved context chunks with unique identifiers. [CITATION_MAP] must link each citation marker to its intended source ID. For high-stakes domains, set [RISK_LEVEL] to "high" and ensure a human reviews any output where overall_drift_score exceeds 0.3 or any individual citation has a drift_distance of "moderate" or "severe". The [EXAMPLES] placeholder should include one clean citation and one drifted citation with the expected correction to calibrate the model's drift distance classification. Always validate the output JSON against the schema before passing it downstream, and log any parse failures for retry with a repair prompt.

The Citation Drift Detection and Correction Prompt is not a standalone artifact; it is a recovery module inside a larger cited-answer pipeline. The typical integration point is after a long-form answer has been generated and before it reaches the user or a downstream audit system. The harness calls the drift detection prompt with the original source passages, the generated answer, and the intended citation anchors. The prompt returns a drift analysis—identifying which citations have shifted away from their source passages—and, optionally, a corrected answer with re-anchored citations. This output must be validated before the corrected answer replaces the original.

Implement the harness as a post-generation step with a clear contract: input includes source_passages (list of objects with passage_id and text), generated_answer (the full text with inline citation markers), and citation_map (mapping each marker to its intended passage_id). The prompt returns a structured JSON object containing drift_detected (boolean), drifted_citations (list of markers with drift_distance scores and explanations), and corrected_answer (the regenerated text with re-anchored citations). Validate the output schema strictly: if drift_distance is missing for any flagged citation, or if the corrected_answer contains citation markers not present in the source passages, reject the output and retry with a more explicit constraint. Log every drift detection result—including drift distances, correction decisions, and retry counts—to your observability platform for trend analysis and model performance tracking.

Model choice matters here. Drift detection requires strong instruction-following and long-context reasoning; use a capable model such as Claude 3.5 Sonnet or GPT-4o for the detection step. For cost-sensitive pipelines, you can route low-risk answers (short responses, single-source answers) through a lighter model and reserve the full drift detection for multi-paragraph, multi-source outputs. Set a retry budget: if the prompt fails validation twice, escalate to a human review queue rather than looping indefinitely. The corrected answer should never be shown to users without passing a final citation completeness check—verify that every claim in the corrected text has a corresponding citation marker and that every marker resolves to a valid passage in the retrieval corpus.

For RAG systems with retrieval retry capability, integrate drift detection with the Evidence Gap Identification and Fill Prompt. When drift is detected because a citation has wandered into territory not covered by its source passage, the harness can trigger targeted re-retrieval to find better evidence for the drifted claim, then regenerate only the affected paragraph rather than the entire answer. This partial regeneration approach reduces latency and preserves the stable portions of the response. Always compare the drift-corrected answer against the original using a citation coverage eval: measure the percentage of factual claims that are properly anchored, and reject corrections that reduce coverage.

Avoid deploying drift detection without human-in-the-loop review for high-stakes domains such as healthcare, legal, or financial compliance. In these contexts, the harness should flag answers with drift_distance scores above a configurable threshold and route them to a review interface where a human can approve the correction, revert to the original, or request manual rewriting. The review interface should display the original answer, the drift analysis, and the corrected answer side by side, with source passages expandable for quick verification. This design keeps the AI system productive while maintaining the evidentiary chain required for audit and compliance.