Verbatim Quote vs Paraphrase Attribution Prompt

This prompt is for RAG system builders and AI engineers who need every statement in a generated answer to be explicitly labeled as either a verbatim quote or a paraphrase. The prompt forces the model to distinguish between exact text pulled from source documents and content it has synthesized, and to attach the correct citation markers to each. Use this when your product requires auditable attribution, when downstream review processes must verify quote accuracy, or when compliance rules mandate clear separation of original source language from AI-generated summary.

This prompt belongs in the answer-generation step of a RAG pipeline, after retrieval and evidence ranking are complete. It assumes you have already selected the most relevant source passages and are now ready to produce a labeled, citable answer. Do not use this prompt when you need fast, low-latency responses without attribution overhead, when your sources are not chunked with stable identifiers, or when the user experience cannot accommodate inline labels. It is also the wrong choice for creative or open-ended generation tasks where source fidelity is not the primary requirement.

Before deploying this prompt, ensure your retrieval pipeline produces passages with unique, stable identifiers that can survive through to the final answer. Test the prompt with edge cases: sources that are nearly identical, passages where the boundary between quote and paraphrase is ambiguous, and inputs where the model might fabricate a quote to satisfy the labeling requirement. Pair this prompt with a post-generation validation step that checks whether every [VERBATIM] label matches the exact text in the cited source. If your use case involves regulated content, add a human review step for any answer where the quote-to-paraphrase ratio exceeds a defined threshold or where the model's confidence in its labeling is low.

The prompt below is designed to be pasted directly into your system message or user message template. It instructs the model to classify every claim it makes about source material into one of two categories—[VERBATIM] for exact quotes and [PARAPHRASED] for synthesized content—and to attach the corresponding source identifier. Replace the square-bracket placeholders with your application's runtime values before sending the request. The template includes explicit rules for quote boundaries, mislabeling prevention, and what to do when the source text is ambiguous.

textCopy
You are an attribution engine. Your job is to answer the user's question using ONLY the provided source documents. You must label every factual statement you make as either a verbatim quote or a paraphrase.

## INPUT

**User Question:** [USER_QUESTION]

**Source Documents:**
[SOURCE_DOCUMENTS]

## RULES

1. **Verbatim Quotes:** If you reproduce text exactly as it appears in a source, wrap the statement in `[VERBATIM]...[END VERBATIM]` tags. Immediately after the closing tag, include the source ID in brackets: `[source: [SOURCE_ID]]`. Do not alter punctuation, spelling, or wording inside a verbatim block.
2. **Paraphrased Statements:** If you restate, summarize, or synthesize information from a source in your own words, prefix the statement with `[PARAPHRASED]` and end it with `[source: [SOURCE_ID]]`. A paraphrase must not contain any sequence of five or more consecutive words identical to the source.
3. **Mixed Attribution:** If a single sentence contains both a verbatim phrase and paraphrased material, break it into separate labeled segments.
4. **No Unlabeled Claims:** Every factual claim that draws on a source document must carry exactly one label and one source ID. Common knowledge or logical deductions that do not depend on the sources should be labeled `[INFERRED]` and must not cite a source.
5. **Ambiguous Attribution:** If you cannot determine whether a statement is verbatim or paraphrased because the source is unclear, label it `[UNCERTAIN]` and include the best available source ID. Add a brief note in parentheses explaining the ambiguity.
6. **Quote Length Limit:** No single verbatim quote may exceed [MAX_QUOTE_LENGTH] words. If a longer quote is required, break it into multiple labeled segments or switch to paraphrase.
7. **Source Conflict:** If two sources contradict each other, present both claims with their labels and source IDs, then add a `[CONFLICT]` note summarizing the disagreement.

## OUTPUT FORMAT

Return your answer as plain text with inline labels. Do not use JSON or markdown tables for the answer body.

## EXAMPLE

**User Question:** What is the capital of France?

**Source Document 1 (ID: src_01):** "Paris is the capital and most populous city of France."

**Answer:**
[VERBATIM]Paris is the capital and most populous city of France.[END VERBATIM][source: src_01]

Adapt the template by adjusting the [MAX_QUOTE_LENGTH] placeholder to match your fair-use or compliance policy. If your source documents carry multiple metadata fields—such as page number, document version, or author—extend the source ID format to [source: [SOURCE_ID], page: [PAGE], date: [DATE]] and update the rules accordingly. For high-stakes domains like legal or clinical review, add a rule that requires human review when [UNCERTAIN] or [CONFLICT] labels appear. Test the prompt with edge cases: sources that are nearly identical, documents with OCR errors, and questions where the answer spans multiple chunks. If the model consistently mislabels short paraphrases as verbatim, tighten the consecutive-word threshold in Rule 2. If it overuses [INFERRED] to avoid citing sources, add a rule that [INFERRED] claims must be explicitly justified in a parenthetical note.

This prompt is designed to sit between your retrieval system and the final answer renderer. It expects a user question and a set of retrieved source documents, and it returns a structured analysis where every factual claim is labeled as either verbatim or paraphrased with a corresponding citation. The implementation harness must enforce that the model's output conforms to this contract before the answer reaches the user. Because mislabeling a paraphrase as a verbatim quote is a factual error that can propagate downstream, the harness should treat label validation as a hard gate, not a best-effort suggestion.

Wire the prompt into your application as a post-retrieval, pre-rendering step. The typical flow: (1) user query triggers retrieval, (2) retrieved chunks are assembled into the [CONTEXT] variable with source IDs, (3) the prompt is invoked with the query and context, (4) the output is parsed and validated, (5) valid outputs proceed to the UI or API response, and (6) invalid outputs trigger a retry or escalate to human review. For validation, implement a post-processing check that extracts every verbatim claim, normalizes whitespace and punctuation, and performs a substring or fuzzy match against the provided source text. If a verbatim claim does not appear in the cited source, flag it as a mislabeling failure and either retry with a stricter temperature setting or route to a review queue. For paraphrased claims, validate that a source citation is present and that the claim does not match the source text verbatim beyond a configurable n-gram overlap threshold (e.g., 5 consecutive words).

Model choice matters here. Use a model with strong instruction-following and low hallucination rates for structured output tasks—GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro are reasonable defaults. Set temperature low (0.0–0.2) to reduce creative rephrasing that could blur the verbatim/paraphrase boundary. Implement a retry strategy: on the first validation failure, resubmit the same prompt with the validation error appended as a [CORRECTION_FEEDBACK] variable. On a second failure, log the full prompt, context, and output for manual review and fall back to a conservative mode that treats all claims as paraphrased with a visible uncertainty flag. For high-stakes domains (legal, clinical, compliance), always route verbatim claims that fail substring verification to a human reviewer before publication. Log every invocation with the prompt version, model, retrieval set IDs, validation results, and any human overrides to maintain an audit trail for citation accuracy over time.