Regulatory Filing Citation Generation Prompt Template

This prompt is designed for compliance engineering teams that need to generate precise, auditable citations to specific locations within regulatory filings. The core job-to-be-done is converting a user query or an extracted fact into a structured citation string that includes the document section, paragraph, page number, and a confidence score. The ideal user is a developer or compliance engineer integrating this prompt into an automated document review pipeline, a RAG system, or an audit evidence packager. Required context includes the full text of the source document with preserved page and paragraph markers, the specific statement or requirement that needs a citation, and a defined output schema for downstream consumption.

Use this prompt when the cost of an incorrect or vague citation is high—such as in SOX audit evidence, SEC filing reviews, or GDPR compliance mapping. It is appropriate when you have already extracted a target claim or obligation and now need to anchor it to a precise source location. The prompt expects pre-processed document text with stable location identifiers; it is not designed to perform the initial extraction or to interpret scanned images without an OCR layer. Do not use this prompt for open-ended document Q&A where the user asks broad questions without a specific target statement. It is also unsuitable for documents where page and paragraph structure cannot be reliably preserved, such as heavily redacted filings or unstructured email threads.

Before wiring this prompt into a production pipeline, ensure your document ingestion process injects and preserves location markers like [PAGE_12] and [PARA_4] into the text. The prompt's value comes from its ability to output structured citation objects with confidence scores, which enables downstream programmatic checks. If a citation falls below your confidence threshold, route it for human review rather than silently accepting it. This prompt is a component in a larger system, not a standalone solution. Pair it with the 'Regulatory Filing Confidence Scoring' and 'Missing Field Detection' playbooks to build a complete, auditable extraction pipeline.

This prompt template is designed to be dropped into your application's prompt assembly layer. It instructs the model to act as a compliance citation specialist, producing structured citation objects that link every claim back to a specific section, paragraph, and page number in the source document. The template uses square-bracket placeholders for all dynamic inputs, making it safe to populate programmatically before each inference call. The core instruction set enforces strict output formatting, abstention when evidence is missing, and explicit confidence scoring for every generated citation.

textCopy
You are a compliance citation specialist. Your task is to generate precise, source-grounded citations from the provided regulatory filing document. You must link every factual claim or extracted data point to its exact location in the source text.

## INPUT
Document Text:
[DOCUMENT_TEXT]

Extraction Target or Question:
[TARGET_QUERY]

## OUTPUT SCHEMA
Return a single JSON object with the following structure. Do not include any text outside the JSON object.
{
  "citations": [
    {
      "citation_string": "string (e.g., 'Section 3.2, Paragraph 4, Page 12')",
      "source_text_excerpt": "string (the verbatim text from the document that supports this citation)",
      "section_reference": "string (the heading or section number)",
      "paragraph_number": "integer or null",
      "page_number": "integer or null",
      "confidence_score": "float between 0.0 and 1.0",
      "grounding_rationale": "string (brief explanation of why this excerpt supports the target query)"
    }
  ],
  "unanswerable_claims": [
    "string (list any parts of the target query that could not be cited from the document)"
  ]
}

## CONSTRAINTS
1. **Source Grounding**: Every citation must include a `source_text_excerpt` that is a verbatim copy from [DOCUMENT_TEXT]. Do not paraphrase.
2. **Abstention**: If a claim in [TARGET_QUERY] cannot be supported by the document, place it in `unanswerable_claims`. Do not invent citations.
3. **Confidence Scoring**: Assign a `confidence_score` based on how directly the excerpt supports the claim. Use 0.9-1.0 for explicit matches, 0.7-0.89 for strong inferences, and below 0.7 for weak or indirect support. If the score is below 0.7, consider adding the claim to `unanswerable_claims`.
4. **Format**: Output only the valid JSON object. No markdown fences, no introductory text.

## RISK LEVEL
[RISK_LEVEL]

## EXAMPLES
[EXAMPLES]

To adapt this template for your specific use case, populate the placeholders before each call. [DOCUMENT_TEXT] should contain the full text of the regulatory filing, pre-processed to remove any extraction artifacts. [TARGET_QUERY] can be a natural language question or a structured list of fields to extract. The [RISK_LEVEL] placeholder should be set to 'high', 'medium', or 'low' to adjust the model's conservatism—for high-risk filings like 10-Ks or drug approval documents, the prompt should be followed by a validation step that checks every citation's source_text_excerpt against the original document using a substring match. The [EXAMPLES] placeholder is critical for calibration; provide at least two few-shot examples that demonstrate the correct handling of explicit matches, weak inferences, and unanswerable claims. After generating citations, always run an evaluation check that verifies: (1) every citation_string resolves to a real location in the document, (2) every source_text_excerpt is a verbatim substring, and (3) no hallucinated sections or page numbers appear in the output.

This prompt is designed to be called after document chunking and retrieval steps, not as a standalone tool. The typical harness flow: (1) retrieve the relevant regulatory filing section, (2) pass the section text and the target claim or question into the prompt, (3) receive a structured citation object, (4) validate the citation against the source document, and (5) log the result for audit. The prompt expects [SOURCE_TEXT] to contain the exact text of the section being cited, not a summary or embedding. If your pipeline uses page-level or paragraph-level chunking, include the chunk's page number and section heading in [CONTEXT] so the model can anchor its citation string.

Validation is the critical harness component. After the model returns a citation, run a grounding check: verify that the quoted or referenced text substring actually appears in [SOURCE_TEXT]. A simple string containment check catches many hallucinations, but fuzzy matching with a configurable threshold (e.g., Levenshtein distance ≤ 5% of the quoted string length) handles OCR noise and minor formatting differences. For the confidence score field, set a minimum threshold (typically 0.85 for regulatory use) below which the citation is routed for human review rather than accepted automatically. Log every citation attempt—including the model, prompt version, source text hash, raw output, validation result, and reviewer action—to an immutable audit table. This traceability is non-negotiable for compliance workflows.

For model choice, prefer models with strong instruction-following and low hallucination rates on extraction tasks. Claude 3.5 Sonnet and GPT-4o are suitable defaults. Avoid smaller or older models that tend to fabricate page numbers or paragraph references when the source text is ambiguous. Implement retry logic with a maximum of two retries: if validation fails on the first attempt, append the specific validation error (e.g., 'Quoted text not found in source') to the retry prompt as [PREVIOUS_ERROR]. If the second attempt also fails, escalate to a human review queue with the full context attached. Do not loop indefinitely—each retry increases latency and cost without guaranteed improvement.

When integrating into a RAG system, ensure the retriever passes full section text rather than top-k snippets. Citation generation degrades sharply when the model only sees fragments. If your retriever cannot guarantee complete section coverage, add a pre-processing step that fetches the full section by its heading or section number before calling this prompt. For multi-section citations, call the prompt once per section and merge results in the application layer rather than asking the model to cite across multiple sections in a single call—this reduces cross-section hallucination. Finally, include a human approval gate for any citation used in regulatory submissions, audit responses, or external communications, regardless of confidence score.