Citation Integrity Verification Prompt After Truncation

This prompt is for RAG system operators and AI engineers who have just compressed or truncated retrieved context and need to verify that every citation marker in the compressed output still maps to a retrievable source passage. The job-to-be-done is preventing orphaned citations—references that point to evidence that was dropped during compression—from reaching end users. If your system generates answers with inline citations (e.g., [1], [doc-4]) and you run a compression step before answer generation, you need this verification gate between compression and the final model call.

Use this prompt when your compression pipeline is lossy—when it summarizes, prunes, or reorders passages and you cannot guarantee that all citation targets survive. It is especially critical in regulated or high-trust domains (legal, clinical, financial) where a citation to missing evidence creates audit risk. Do not use this prompt if your compression is purely extractive and preserves every source passage intact with its original identifier; in that case, citation integrity is structural, not semantic. Also skip this prompt if your downstream answer generator does not produce citations at all—there is nothing to verify.

The ideal user is an engineer building a production RAG pipeline who needs an automated verification step that runs after compression and before answer generation. You should wire this prompt into a pre-inference validation harness: if the verification report flags orphaned citations, either repair the compressed context by re-inserting the missing passages, drop the orphaned citation markers, or escalate for human review depending on your risk tolerance. For high-stakes workflows, always require human approval when orphaned citations are detected rather than silently dropping them.

This prompt template is designed to be dropped into a RAG pipeline immediately after a context compression or truncation step. Its job is to verify that every citation marker present in the compressed context still references a passage that exists in the original source material and that no orphaned citations—markers pointing to dropped or hallucinated sources—remain. The template uses square-bracket placeholders so you can substitute your own compressed context, original source passages, citation format rules, and risk thresholds without rewriting the core verification logic.

textCopy
You are a citation integrity verifier for a RAG system. Your task is to confirm that every citation marker in the [COMPRESSED_CONTEXT] maps to a retrievable passage in the [ORIGINAL_SOURCES] and that no orphaned citations exist.

## INPUTS
- Compressed context: [COMPRESSED_CONTEXT]
- Original source passages (with source IDs): [ORIGINAL_SOURCES]
- Citation format: [CITATION_FORMAT] (e.g., "[1]", "(Author, Year)", "§4.2")
- Risk level: [RISK_LEVEL] (low, medium, high, critical)

## OUTPUT SCHEMA
Return a JSON object with the following structure:
{
  "verification_report": {
    "total_citations_found": <integer>,
    "verified_citations": <integer>,
    "orphaned_citations": <integer>,
    "citation_details": [
      {
        "citation_marker": "<string>",
        "status": "verified | orphaned | ambiguous",
        "mapped_source_id": "<string or null>",
        "mapped_passage_excerpt": "<string or null>",
        "confidence": "high | medium | low",
        "issue_description": "<string or null>"
      }
    ],
    "overall_pass": <boolean>,
    "critical_failures": ["<string>"]
  }
}

## CONSTRAINTS
1. Do not infer or fabricate source passages. If a citation marker cannot be matched to an exact passage in [ORIGINAL_SOURCES], mark it as orphaned.
2. For ambiguous matches where a citation could refer to multiple passages, mark status as "ambiguous" and list all candidate source IDs in issue_description.
3. If [RISK_LEVEL] is "high" or "critical", flag any orphaned or ambiguous citation as a critical_failure and set overall_pass to false.
4. If [RISK_LEVEL] is "low", orphaned citations with low confidence may still pass if they represent less than 5% of total citations.
5. Preserve the exact citation marker string as it appears in [COMPRESSED_CONTEXT] in the citation_marker field.
6. If [COMPRESSED_CONTEXT] contains no citation markers at all, return total_citations_found: 0, overall_pass: true, and note this in critical_failures only if citations were expected per [CITATION_FORMAT].

## VERIFICATION STEPS
1. Extract every citation marker from [COMPRESSED_CONTEXT] using the [CITATION_FORMAT] pattern.
2. For each marker, search [ORIGINAL_SOURCES] for the corresponding source ID and passage.
3. Confirm that the cited claim or fact in the compressed context is supported by the mapped passage.
4. If a marker references a source ID not present in [ORIGINAL_SOURCES], mark as orphaned.
5. If a marker references a valid source ID but the passage content does not support the surrounding claim, mark as ambiguous with an explanation.
6. Compile the verification report and set overall_pass based on [RISK_LEVEL] thresholds.

To adapt this template, replace each square-bracket placeholder with your pipeline's actual values. For [ORIGINAL_SOURCES], pass the full source passages with their IDs as they existed before compression—not the compressed versions. For [CITATION_FORMAT], provide a regex-ready pattern or a natural-language description of your citation style. The [RISK_LEVEL] placeholder controls the strictness of the pass/fail decision; wire this to your application's risk configuration so that regulated or customer-facing workflows fail closed on any orphaned citation. After the verification report is returned, your application layer should gate the downstream use of the compressed context: if overall_pass is false, either repair the context, re-retrieve dropped sources, or escalate for human review before the compressed context reaches the answer-generation step.

The citation integrity verification prompt is designed to run as a post-truncation validation step in a RAG pipeline, not as a standalone tool. After your context compression or truncation module produces a condensed context with embedded citation markers, this prompt acts as a quality gate before the final generation step. The typical integration point is between the context assembly layer and the answer generation model: the compressed context and its source mapping are passed to this verification prompt, and only contexts that pass verification proceed to the answer generator. This prevents orphaned citations—markers that reference source passages no longer present in the compressed context—from reaching the user as broken or unverifiable claims.

To wire this into an application, build a verification function that accepts two inputs: the compressed context string containing citation markers (e.g., [1], [2]) and a source mapping object that links each marker to its original retrievable passage. The function should call the verification prompt with both inputs and parse the structured output—a JSON object containing a verification_status field (pass or fail), a citation_map array listing each marker and whether it maps to a present passage, and an orphaned_citations array listing any markers without matching source content. Implement a retry loop: if verification fails due to orphaned citations, route the failure back to your truncation module for correction before re-verification. Set a maximum of three retry cycles before escalating to a human reviewer or logging the failure for offline analysis. For model choice, use a model with strong instruction-following and structured output capabilities—GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro are suitable. Avoid smaller or older models that may hallucinate verification results, defeating the purpose of the check.

Validation should include both structural and semantic checks. Structurally, confirm that the verification output is valid JSON matching your expected schema and that every citation marker from the compressed context appears in the citation_map. Semantically, implement a spot-checking mechanism: randomly sample 10-20% of verified citation-to-passage mappings and run a secondary factuality check comparing the cited passage against the compressed context's claim. This catches cases where the verification model incorrectly passes a citation that references distorted or summarized content rather than the original passage. Log every verification result with the compressed context, source mapping, verification output, and any spot-check discrepancies. For high-stakes domains such as legal, medical, or financial applications, add a human review gate that requires manual approval for any verification run where orphaned citations were detected and repaired, or where spot-checking reveals semantic drift between the compressed context and source passages. Do not allow auto-repaired contexts to reach end users in regulated workflows without human sign-off.

Common failure modes in the harness include: the verification model hallucinating citation markers that don't exist in the compressed context, producing malformed JSON under high token pressure, or passing citations where the compressed context paraphrased the source beyond recognition. Mitigate these by setting a low temperature (0.0-0.1) for deterministic outputs, using structured output APIs where available, and implementing a schema validator that rejects any verification response missing required fields. If the verification model consistently fails on long compressed contexts, split the context into chunks of 3-5 citation markers each and run verification in parallel, then merge results. Monitor verification latency and cost—this prompt adds an extra inference call to your pipeline, so track its token usage and set alerts if verification costs exceed 20% of your total RAG pipeline spend. The verification step should be fast and cheap relative to the generation step it protects.