Citation Span Extraction for RAG Answers Prompt

This prompt is for RAG system builders who need every sentence in a generated answer to be anchored to a specific, verifiable span in the retrieved context. It forces the model to extract the exact text that supports each claim rather than pointing to an entire document or passage. The primary job-to-be-done is audit-grade traceability: a downstream reviewer or automated system must be able to click from a claim directly to the source text that proves it, without hunting through a multi-paragraph block. Use this when answers will be reviewed for factual accuracy, when you need to detect hallucinated claims that lack source support, or when your compliance framework requires granular provenance for every generated statement.

This prompt is not a replacement for retrieval quality. It assumes the relevant evidence is already present in the context window. If retrieval fails to surface the right documents, no citation extraction can fix the gap. The ideal user is an engineering lead or AI architect integrating a RAG pipeline into a product where trust depends on verifiability—think legal research tools, clinical decision support summaries, financial audit assistants, or any system where a wrong answer with a confident citation is worse than no answer at all. Before using this prompt, confirm that your retrieval step reliably returns the source material needed to answer the question. If your retrieval recall is below 90% for the target query distribution, invest there first.

Do not use this prompt when the cost of extraction outweighs the value of granular traceability. For consumer-facing chatbots where answers are short and the risk of hallucination is managed through other means, a simpler citation approach—such as linking to a whole document or passage—may be sufficient. This prompt adds latency and token overhead because it requires the model to copy exact source spans into the output. It also introduces a new failure mode: over-citation, where the model attaches a source to every sentence even when the connection is weak, creating a false sense of rigor. Reserve this prompt for workflows where the cost of a missed hallucination is high and the downstream consumer has the capability to act on span-level citations.

The prompt below is designed to run after your RAG pipeline has already produced an answer from retrieved documents. It takes the generated answer and the retrieved context as inputs, then outputs a structured mapping of each answer sentence to the specific text spans that support it. This is not a generation prompt—it is a post-hoc extraction and verification step that makes citation traceable before the answer reaches the user or an auditor.

textCopy
You are a citation extraction system. Your task is to map every sentence in the provided answer to the exact source spans in the provided context that support it.

## INPUT

### ANSWER
[ANSWER_TEXT]

### RETRIEVED CONTEXT
[CONTEXT_BLOCKS]

## OUTPUT SCHEMA
Return a JSON object with the following structure:
{
  "citations": [
    {
      "answer_sentence_index": 0,
      "answer_sentence_text": "string",
      "source_spans": [
        {
          "context_block_id": "string",
          "span_text": "string",
          "span_start_char": 0,
          "span_end_char": 0,
          "support_type": "direct_quote | paraphrase | inference",
          "confidence": 0.0
        }
      ],
      "is_unsupported": false,
      "unsupported_reason": null
    }
  ],
  "unsupported_sentences": [
    {
      "answer_sentence_index": 0,
      "answer_sentence_text": "string",
      "reason": "no_relevant_context | context_contradicts | insufficient_evidence"
    }
  ],
  "over_citation_warnings": [
    {
      "answer_sentence_index": 0,
      "warning": "string",
      "suggested_span_reduction": ["context_block_id"]
    }
  ]
}

## CONSTRAINTS
- Map every answer sentence, including introductory and transitional sentences.
- For each source span, include the exact character offsets from the context block.
- If a sentence is fully supported by a direct quote, mark support_type as "direct_quote".
- If a sentence paraphrases source material, mark support_type as "paraphrase".
- If a sentence draws an inference that goes beyond the source text, mark support_type as "inference" and reduce confidence accordingly.
- If a sentence has no supporting evidence in the context, add it to unsupported_sentences with a specific reason.
- Flag over-citation when more than three source spans are cited for a single sentence and at least one is redundant.
- Do not cite spans that are tangentially related but do not directly support the claim.
- If the context contains contradictory information, note it in the unsupported reason.
- Confidence must be between 0.0 and 1.0, where 1.0 is a verbatim match and lower values reflect paraphrasing or inference.

## EXAMPLES

### Example 1: Direct Quote
Answer sentence: "The company reported revenue of $4.2 billion in Q3."
Context: "Acme Corp announced Q3 revenue of $4.2 billion, up 12% year-over-year."
Output:
{
  "answer_sentence_index": 0,
  "answer_sentence_text": "The company reported revenue of $4.2 billion in Q3.",
  "source_spans": [
    {
      "context_block_id": "block_1",
      "span_text": "Q3 revenue of $4.2 billion",
      "span_start_char": 24,
      "span_end_char": 50,
      "support_type": "direct_quote",
      "confidence": 0.98
    }
  ],
  "is_unsupported": false,
  "unsupported_reason": null
}

### Example 2: Unsupported Claim
Answer sentence: "The CEO attributed the growth to the new AI product line."
Context: "Acme Corp announced Q3 revenue of $4.2 billion, up 12% year-over-year. The company launched three new products this year."
Output:
{
  "answer_sentence_index": 1,
  "answer_sentence_text": "The CEO attributed the growth to the new AI product line.",
  "source_spans": [],
  "is_unsupported": true,
  "unsupported_reason": "no_relevant_context"
}

## RISK LEVEL
HIGH — Incorrect citations can create false audit trails and mislead users about source fidelity. Every output must be validated before production use.

Adapt this template by replacing [ANSWER_TEXT] with the full RAG-generated answer and [CONTEXT_BLOCKS] with your retrieved passages, each tagged with a unique context_block_id. If your context blocks already include character offsets from source documents, pass them through unchanged. If not, pre-process your context to add block IDs and character indexing before calling this prompt. The output schema is designed to feed directly into a citation validator or a human review queue—do not display unsupported sentences to end users without first flagging them.

Before deploying, run this prompt against a golden dataset of 50–100 answer-context pairs with known citation ground truth. Measure citation recall (did we cite every supported sentence?), citation precision (are cited spans actually supportive?), and over-citation rate (flagged warnings per answer). If over-citation exceeds 15% or unsupported sentence detection misses more than 5% of known unsupported claims, adjust the constraints or add few-shot examples before shipping.

The citation span extraction prompt is not a standalone tool; it is a post-retrieval verification step that sits between your RAG generator and the final answer payload. In a production harness, you first run your standard RAG pipeline to produce an answer with inline citation markers. Then, for each answer sentence and its claimed source, you call this prompt with the sentence and the full retrieved context chunk. The prompt returns a minimal, verifiable citation span—or a null result if the sentence is not supported. This two-stage design keeps the generator fast while adding a focused verification layer that catches hallucinated citations before they reach the user.

Validation and retry logic must be strict. After receiving the model output, validate that the returned citation_span is a substring of the provided [CONTEXT]. If it is not, retry once with an explicit instruction that the span must be an exact substring. If the retry also fails, log the failure and flag the sentence for human review. For high-compliance use cases, implement a second validation pass that checks whether the citation span semantically supports the claim using a lightweight NLI model or an LLM judge prompt. Sentences that fail both substring and semantic validation should never be presented as cited facts. Model choice matters: use a model with strong instruction-following and low hallucination rates for this task—GPT-4o or Claude 3.5 Sonnet are good defaults. Avoid smaller or older models that may paraphrase the span instead of extracting it verbatim.

Observability and logging are critical because citation failures are often the first signal of RAG pipeline degradation. Log every extraction attempt with the input sentence, the retrieved context, the extracted span, the substring validation result, and the retry count. Track metrics like citation extraction latency, substring match rate, retry rate, and the percentage of sentences that receive a null citation. Set alerts on sudden drops in the match rate, which often indicate a change in the retriever, chunking strategy, or generator prompt. Store a sample of failed extractions for weekly review. Human review integration should be designed as a queue, not a blocker: sentences that fail validation after retries are routed to a review interface where a human can confirm, correct, or reject the citation. The corrected spans should be fed back into your evaluation dataset to improve future prompt versions. Avoid wiring this prompt directly into user-facing latency budgets—run it asynchronously after the initial answer is streamed, and update citations in the UI once verification completes.