Source-Anchored Quote Extraction Prompt

This prompt is for engineering teams building audit trails, compliance systems, and high-stakes AI applications where every extracted piece of evidence must carry a complete, verifiable provenance chain. The job-to-be-done is not just finding a relevant quote, but producing a structured record that links a verbatim text span to its source document ID, author, date, and retrieval rank. The ideal user is a developer or AI architect who needs to wire this extraction into a downstream system of record—such as an audit log, a regulatory filing, or a legal review tool—where a quote without metadata is a liability.

Use this prompt when the cost of a missing or unverifiable source is high. It is designed for scenarios where you must answer the question 'Where exactly did this come from, and why should we trust it?' before any human reviewer sees the output. This is the right choice when you are building for legal-tech, financial compliance, medical documentation, or any regulated workflow that demands source anchoring as a non-negotiable output contract. The prompt enforces a strict schema that rejects any extracted span lacking a document ID, author, or date, making it suitable for automated pipelines that cannot tolerate incomplete provenance.

Do not use this prompt for casual Q&A, creative writing, or low-stakes summarization where source metadata is irrelevant. It is over-engineered for simple citation tasks that only need a quote and a document title. If your application only requires inline citations for a chat interface, start with the simpler Inline Citation Span Selection Prompt or RAG Citation Answer Prompt Template instead. This prompt is also a poor fit when your source documents lack structured metadata—if you cannot provide document IDs, authors, or dates, the mandatory schema will cause valid quotes to be rejected, and you will get empty outputs instead of useful evidence. Before adopting this prompt, ensure your retrieval pipeline can supply the required metadata fields for every candidate document.

This template enforces source anchoring for every extracted quote by requiring document ID, date, author, and retrieval rank in the output schema. It is designed for audit trail and compliance system builders who need complete provenance chains, not just relevant text spans. The prompt separates extraction instructions from output formatting rules so you can adapt the extraction logic without breaking downstream validation.

codeCopy
You are an evidence extraction specialist. Your task is to extract verbatim quotes from source documents that directly support a given claim. Every quote must be anchored to its source with complete provenance metadata.

## INPUT

**Claim:** [CLAIM]

**Source Documents:**
[SOURCE_DOCUMENTS]

## EXTRACTION RULES

1. Extract only verbatim text spans. Do not paraphrase, summarize, or alter the original text.
2. Each quote must directly support or refute the claim. Do not extract tangentially related passages.
3. Prefer the most specific and concise span that captures the supporting evidence. Avoid extracting entire paragraphs when a sentence suffices.
4. If multiple documents contain relevant evidence, extract quotes from each, up to a maximum of [MAX_QUOTES_PER_DOCUMENT] quotes per document.
5. If no document contains supporting evidence, return an empty quotes array and set `evidence_found` to false.
6. Preserve original formatting, including line breaks within quotes, but trim leading and trailing whitespace.
7. For quotes that reference defined terms, figures, or cross-references, include enough surrounding context to make the quote interpretable without the full document.

## OUTPUT SCHEMA

Return a JSON object with this exact structure:

```json
{
  "claim": "[CLAIM]",
  "evidence_found": true,
  "quotes": [
    {
      "quote_text": "Verbatim extracted text",
      "source_document_id": "[DOCUMENT_ID]",
      "source_document_title": "[DOCUMENT_TITLE]",
      "source_date": "[DOCUMENT_DATE]",
      "source_author": "[DOCUMENT_AUTHOR]",
      "retrieval_rank": [RETRIEVAL_RANK],
      "span_location": {
        "start_offset": [START_OFFSET],
        "end_offset": [END_OFFSET],
        "section": "[SECTION_HEADING_OR_NULL]"
      },
      "relevance_rationale": "Brief explanation of how this quote supports or refutes the claim"
    }
  ],
  "extraction_notes": "[OPTIONAL_NOTES_ABOUT_AMBIGUITY_OR_CONFLICTING_EVIDENCE]"
}

CONSTRAINTS

• Do not fabricate source metadata. If a field is unavailable, use null.
• Do not extract the same quote more than once.
• If two sources contain near-identical quotes, extract both and note the duplication in extraction_notes.
• For time-sensitive claims, prefer the most recent source when multiple sources provide equivalent support.
• If the claim contains multiple sub-claims, extract quotes for each sub-claim and note which sub-claim each quote supports in relevance_rationale.

Adaptation guidance: Replace [CLAIM] with the statement requiring evidence support. Populate [SOURCE_DOCUMENTS] with pre-retrieved passages formatted as document objects containing ID, title, date, author, retrieval rank, and full text. Set [MAX_QUOTES_PER_DOCUMENT] based on your application's citation density requirements—lower values produce cleaner outputs for inline citations, higher values support comprehensive audit trails. The span offsets assume your source documents expose character-level indexing; if your retrieval pipeline does not preserve offsets, remove the span_location field from the output schema and the extraction rules that reference it.

Validation and risk controls: Before ingesting the model's output into an audit system, validate that every source_document_id matches a known document in your retrieval set, that quote_text is a substring of the corresponding source document text, and that retrieval_rank values are consistent with your retrieval order. For compliance workflows, route outputs with evidence_found: false or conflicting quotes to a human reviewer. Run periodic evals comparing extracted quotes against human-annotated evidence spans to measure precision and recall drift across prompt versions.

The Source-Anchored Quote Extraction Prompt is designed to be called as a single step within a larger retrieval and verification pipeline. The typical flow is: retrieve candidate passages, call this prompt to extract and anchor quotes, validate the output schema, and then pass the structured quotes to downstream citation rendering or audit logging. Because every extracted span must carry complete source metadata (document ID, date, author, retrieval rank), the harness must enforce that no quote enters your system without a full provenance chain. This prompt is not a standalone chatbot interaction—it is a structured extraction step that expects pre-retrieved context and returns a strict JSON schema.

Wire the prompt into your application as a stateless function call. The inputs are a [CLAIM] to support or refute and a [SOURCES] array where each source object contains doc_id, date, author, retrieval_rank, and text. The model must return a JSON array of quote objects, each with quote_text, source_doc_id, source_date, source_author, source_retrieval_rank, char_start, char_end, and relevance_rationale. Before any quote reaches downstream consumers, run a post-extraction validator that checks: (1) every source_doc_id matches an input source, (2) quote_text is a verbatim substring of the referenced source text at the specified character offsets, (3) no quote object has null or missing provenance fields, and (4) the output is valid JSON. If validation fails, retry once with the same prompt plus the validation error message appended as a [CONSTRAINTS] update. If the retry also fails, log the failure and route to a human review queue rather than silently dropping malformed extractions.

For model choice, prefer models with strong instruction-following and JSON mode support (GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro with response schema enabled). Set temperature to 0 or near-zero to maximize extraction consistency. If your retrieval pipeline returns more than 8-10 source passages, chunk the sources into batches and run parallel extraction calls, then deduplicate quotes by exact text match and keep the highest-ranked source for each unique quote span. Log every extraction call with the input claim, source IDs, extracted quote count, validation pass/fail, and latency. This log becomes your audit trail for debugging extraction quality over time. Do not use this prompt for open-ended summarization or for claims where no source documents exist—it will hallucinate source metadata if forced to extract from empty context.

When integrating into compliance or audit systems, store the full prompt request and response payloads immutably. The source anchoring metadata (doc_id, date, author, retrieval_rank) must flow through to any downstream citation UI or audit report without modification. If your application displays quotes to users, always render the source metadata alongside the quote text so users can verify provenance themselves. For high-stakes domains (legal, regulatory, clinical), add a human review step before extracted quotes are published or used in decisions. The harness should flag any extraction where relevance_rationale indicates low confidence or where the quote count is zero despite available sources—these are signals that the retrieval set may be insufficient or the claim may be unanswerable from the provided evidence.