Answer-to-Source Span Alignment Prompt Template

This prompt is for builders who need to prove exactly which evidence supports which part of an answer, not just that evidence exists somewhere in the retrieved set. The core job is fine-grained attribution: mapping each sentence or factual claim in a generated answer to the specific source passage, paragraph, or line that supports it, and returning alignment pairs with confidence scores. The ideal user is an AI engineer or product developer integrating citation-aware RAG into a legal-tech product, clinical decision support tool, financial research assistant, or any application where users must verify claims by clicking through to the original source. You need the full generated answer and the complete set of retrieved source documents as inputs—partial context or truncated sources will produce unreliable alignments.

Use this prompt when downstream features depend on inline source linking, citation formatting, or automated grounding audits. It is a core building block for systems that render clickable citations next to each sentence, highlight supported vs. unsupported spans in a UI, or run batch faithfulness checks before publishing AI-generated content. The alignment pairs this prompt produces can feed directly into a citation formatter, a grounding verification pipeline, or a human review queue. You should also use this when you need to debug hallucination at the sentence level—seeing which claims failed to align with any source passage is often more actionable than a single overall faithfulness score.

Do not use this prompt when you only need a passage-level relevance score or a binary grounded/ungrounded verdict. If your application just needs to know whether an answer is generally supported, the Claim-by-Claim Faithfulness Audit or Hallucination Risk Scoring prompts are lighter-weight and faster. Avoid this prompt when the source documents are extremely long and cannot be chunked with stable identifiers—span alignment requires that every passage has a durable reference (section number, paragraph ID, or line range) that survives chunking and retrieval. Also avoid it when latency is the primary constraint and users do not need sentence-level citations; the alignment task is more expensive than a single faithfulness classification. Finally, do not use this prompt as a substitute for human review in regulated domains—it produces alignment evidence for a reviewer to verify, not a final determination of accuracy.

This template performs sentence-level alignment between a generated answer and the provided source passages. It expects the answer to be broken into discrete sentences or claims and returns a structured mapping that identifies which source span supports each unit. The output is designed to feed directly into citation formatting, grounding verification, or audit pipelines. Before using this template, ensure your answer has been segmented—if you need claim-level or clause-level granularity, adjust the instruction in the [GRANULARITY] placeholder accordingly.

codeCopy
You are an alignment specialist. Your task is to map each sentence in the provided answer to the specific source passage, paragraph, or line that supports it.

[GRANULARITY]
Align at the sentence level. Treat each sentence as an independent unit. If a sentence combines information from multiple sources, flag it and note all contributing sources.

[INPUT]
Answer:
"""
[ANSWER_TEXT]
"""

Source passages:
"""
[SOURCE_PASSAGES]
"""

[OUTPUT_SCHEMA]
Return a JSON object with the following structure:
{
  "alignments": [
    {
      "answer_sentence_index": 0,
      "answer_sentence_text": "string",
      "source_passage_id": "string or null",
      "source_span": "exact quoted text from source or null",
      "confidence": 0.0 to 1.0,
      "alignment_type": "direct_match | partial_match | inferred | multi_source | unsupported",
      "notes": "brief explanation if confidence < 0.8 or alignment_type is not direct_match"
    }
  ],
  "coverage_summary": {
    "total_sentences": 0,
    "fully_supported": 0,
    "partially_supported": 0,
    "unsupported": 0,
    "multi_source": 0
  }
}

[CONSTRAINTS]
- Every sentence in the answer must appear in the alignments array.
- Use the exact source text as the source_span. Do not paraphrase.
- If no source supports a sentence, set source_passage_id and source_span to null, confidence to 0.0, and alignment_type to "unsupported".
- If multiple sources are required, set alignment_type to "multi_source" and list all contributing passage IDs in notes.
- Confidence reflects how directly the source span supports the sentence, not how confident the source itself is.
- Do not invent source spans. If you cannot find a match, mark it unsupported.

[EXAMPLES]
Answer sentence: "The company reported $2.3B in revenue for Q3 2024."
Source passage: "Acme Corp announced third-quarter revenue of $2.3 billion, up 12% year-over-year."
Output:
{
  "answer_sentence_index": 0,
  "answer_sentence_text": "The company reported $2.3B in revenue for Q3 2024.",
  "source_passage_id": "doc_3_para_2",
  "source_span": "Acme Corp announced third-quarter revenue of $2.3 billion, up 12% year-over-year.",
  "confidence": 0.95,
  "alignment_type": "direct_match",
  "notes": ""
}

Answer sentence: "This growth was driven by new product launches."
Source passages: No passage mentions product launches.
Output:
{
  "answer_sentence_index": 1,
  "answer_sentence_text": "This growth was driven by new product launches.",
  "source_passage_id": null,
  "source_span": null,
  "confidence": 0.0,
  "alignment_type": "unsupported",
  "notes": "No source passage attributes growth to product launches."
}

[RISK_LEVEL]
High. Incorrect alignments can create false confidence in unsupported claims. Always run alignment outputs through a secondary verification step before surfacing citations to users.

Adapt this template by adjusting the [GRANULARITY] instruction for your use case. For legal or compliance workflows, switch to claim-level alignment where each factual assertion is treated independently. For long-form answers, consider adding a [MAX_SENTENCES] constraint to prevent runaway alignment tasks. The [OUTPUT_SCHEMA] can be extended with additional fields such as source_document_title, source_timestamp, or review_required if your downstream citation formatter needs them. If your source passages lack stable IDs, generate them from document names and paragraph indices before calling this prompt.

After copying this template, wire it into a validation harness that checks every alignment entry has a valid answer_sentence_index, that coverage_summary counts match the alignments array length, and that any confidence above 0.7 has a non-null source_span. For high-risk domains, route alignments with alignment_type of unsupported or inferred to human review before they reach end users. Run this prompt against a golden dataset of known answer-source pairs to calibrate your confidence thresholds before production deployment.

The Answer-to-Source Span Alignment prompt template is designed to be called after answer generation and before the response is delivered to the user. In a typical RAG pipeline, you already have a generated answer and a set of retrieved source passages. This prompt takes both as input and returns a structured mapping of each sentence or claim to the specific source span that supports it. Wire this as a post-processing step: your application calls the LLM to generate the answer, then immediately calls this alignment prompt with the same answer and evidence set. The output is a list of alignment pairs—each containing the answer span, the source span, and a confidence score—that your application can use to format inline citations, build a source panel, or flag unsupported claims for human review.

Integration pattern. Place this prompt inside a dedicated alignment service or function that accepts answer_text, source_passages (as a list of objects with source_id, text, and optional metadata), and alignment_threshold (default 0.7). The function should: (1) assemble the prompt with the answer and passages injected into the [ANSWER] and [SOURCES] placeholders; (2) call your model with response_format set to JSON if using a provider that supports structured output, or append a strict JSON schema instruction to the prompt; (3) parse the response and validate that every alignment object contains the required fields (answer_span, source_id, source_span, confidence); (4) filter out alignments below the threshold; and (5) return the validated list. If the model returns malformed JSON, retry once with a repair prompt that includes the raw output and the expected schema. If the second attempt fails, log the failure and fall back to delivering the answer without inline citations rather than blocking the response.

Model choice and latency. This is a structured extraction task that benefits from models with strong instruction-following and JSON output discipline. GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro all perform well. For lower latency, consider using a smaller model (GPT-4o-mini, Claude 3 Haiku) with a stricter schema constraint and a higher alignment threshold to discard low-confidence pairs. If your application requires real-time citation rendering, run this prompt in parallel with answer generation by passing the answer to the alignment prompt as soon as the first token stream completes. Cache alignment results keyed by (answer_hash, source_set_hash) when the same answer-evidence pair is likely to repeat—for example, in FAQ-style systems or static knowledge bases.

Downstream citation formatting. The alignment output is not the final citation string. Your application must map each source_id to its display metadata (title, URL, page number, section heading) and format the citation according to your product's style. For inline citations, iterate through the answer text and insert superscript markers or tooltip anchors at each answer_span position. For a source panel, group alignments by source_id and display the supporting quotes. Critical check: before rendering, verify that the source_span text actually appears in the referenced source passage. A hallucinated source span that doesn't match the passage text is a silent failure mode—add a string-containment check and discard any alignment where the source span is not a substring of the passage. Log these discards for monitoring.

Evaluation and monitoring. Track three metrics in production: (1) alignment coverage—the percentage of answer sentences that received at least one alignment above threshold; (2) alignment rejection rate—the percentage of alignments discarded due to source-span mismatch or low confidence; and (3) end-to-end latency added by the alignment step. Set up a weekly manual review of 20–30 aligned answers where a reviewer checks whether the source spans genuinely support the answer claims. Use these reviews to calibrate your alignment_threshold. If coverage drops below 80% or rejection rate spikes above 15%, investigate whether your retrieval pipeline is returning insufficient or irrelevant passages. The alignment prompt is a diagnostic signal for your RAG quality, not just a citation formatter.