Missing Citation Flagging and Abstention Prompt Template

This prompt is designed for production RAG systems, document Q&A applications, and compliance tools where a fabricated or missing citation is a higher-risk failure than a refused answer. The core job-to-be-done is enforcing evidence-bound generation: the model must either provide a claim with a verifiable source or explicitly mark the claim as unsupported and abstain from answering. The ideal user is an engineering lead or AI builder integrating a model into a regulated workflow—such as legal contract review, financial audit support, or clinical documentation—where an auditor or end user must be able to trace every factual statement back to a source chunk. The required context includes retrieved document passages with stable identifiers, a defined output schema that separates supported claims from unsupported ones, and a clear abstention policy that tells the model when to stop rather than guess.

Do not use this prompt as a general-purpose Q&A template. It is intentionally restrictive and will refuse to answer questions that a standard RAG prompt might handle by paraphrasing without citation. This trade-off is acceptable when the cost of a hallucinated citation—such as a fabricated case law reference, a wrong financial figure attributed to a filing, or a clinical claim linked to the wrong patient note—is higher than the cost of saying 'I don't have enough evidence.' The prompt is also inappropriate for creative or open-ended tasks where source grounding is not expected. Before deploying, define the minimum evidence threshold for a claim to be considered supported, and ensure your retrieval pipeline passes source metadata (document ID, chunk index, page number) that the model can reference in its output. Without stable source identifiers, the abstention logic will fail because the model cannot reliably distinguish between 'evidence exists but I can't cite it' and 'no evidence exists.'

Next, pair this prompt with an evaluation harness that tests both false abstention (the model refuses to answer when evidence is present) and silent citation gaps (the model answers without marking unsupported claims). Run these evals against a golden dataset where you know exactly which claims should be supported and which should trigger abstention. In high-risk domains, add a human review step for any output that contains abstention markers before it reaches the end user, and log every abstention event for audit. If your system cannot tolerate any abstention, this prompt is the wrong choice—consider instead a prompt that ranks evidence quality and includes confidence scores, allowing downstream code to decide what to surface.

This prompt template is designed for document Q&A and RAG systems where fabricating a citation is worse than admitting ignorance. It instructs the model to answer only when source evidence directly supports a claim, to cite that evidence explicitly, and to produce a structured abstention marker when no supporting source is found. The template uses square-bracket placeholders for your application's specific inputs, output schema, and risk tolerance.

textCopy
You are a citation-verification assistant operating in a high-accuracy document Q&A system. Your primary directive is to never fabricate a citation or claim support where none exists.

## INPUT
User Question: [USER_QUESTION]
Retrieved Source Chunks (each with a unique `chunk_id`):
[RETRIEVED_CHUNKS]

## OUTPUT SCHEMA
Respond with a single JSON object matching this schema:
{
  "status": "answered" | "abstained" | "partial",
  "answer": string | null,  // The answer text if status is "answered" or "partial"; null if "abstained"
  "citations": [
    {
      "chunk_id": string,
      "excerpt": string,  // The exact text from the chunk that supports the claim
      "confidence": "high" | "medium" | "low"
    }
  ],
  "missing_evidence": [string] | null,  // List of specific claims or facts that could not be sourced; null if none
  "abstention_reason": string | null  // Required if status is "abstained" or "partial"
}

## CONSTRAINTS
1. **No Fabrication**: If no retrieved chunk supports a factual claim, do not include that claim in the answer.
2. **Explicit Abstention**: If zero chunks support any part of the question, set `status` to "abstained", `answer` to null, `citations` to an empty array, and provide a clear `abstention_reason`.
3. **Partial Answers**: If some parts of the question are supported and others are not, set `status` to "partial", answer only the supported parts, and list the unsupported claims in `missing_evidence`.
4. **Exact Excerpts**: The `excerpt` field must contain a verbatim string from the source chunk, not a paraphrase.
5. **Confidence Calibration**:
   - `high`: The excerpt directly and unambiguously answers the question.
   - `medium`: The excerpt implies or strongly suggests the answer but requires minor inference.
   - `low`: The excerpt is tangentially related; flag for human review.
6. **Handle Conflicts**: If two chunks provide contradictory evidence, do not pick a winner. Set `status` to "partial", include both citations, and note the conflict in `missing_evidence`.

## EXAMPLES
[FEW_SHOT_EXAMPLES]

## RISK LEVEL
[RISK_LEVEL]  // "high" triggers additional refusal for ambiguous low-confidence cases; "standard" allows medium-confidence answers with flags

To adapt this template, replace [USER_QUESTION] with the user's query and [RETRIEVED_CHUNKS] with your retrieval pipeline's output, ensuring each chunk has a unique chunk_id. Populate [FEW_SHOT_EXAMPLES] with 2-3 demonstrations showing correct abstention, partial answers, and full answers. Set [RISK_LEVEL] to "high" for regulated domains like legal or healthcare where any hallucination is unacceptable, or "standard" for internal analytics where medium-confidence answers are tolerable with a flag. Before deploying, run this prompt against a golden dataset containing questions with known missing evidence to verify that the abstention rate matches your expectations and that no silent citation gaps occur.

This prompt is designed as a post-retrieval safety layer, not a standalone Q&A system. It sits between your retrieval engine and the final answer renderer. The prompt receives retrieved chunks, the user query, and a risk level, then produces a structured output that either cites evidence or explicitly abstains. The core integration pattern is: retrieve → cite-or-abstain → validate → route. Do not send raw model output directly to users without validation, because silent citation gaps are the primary failure mode this prompt exists to prevent.

Integration steps: (1) Retrieve candidate chunks using your existing RAG pipeline. (2) Populate the [RETRIEVED_CHUNKS] placeholder with chunk text and metadata (source ID, page, chunk index). (3) Set [RISK_LEVEL] to high for regulated domains (legal, medical, financial) to enforce strict abstention; use moderate for internal tools where partial answers with flags are acceptable. (4) Call the model with response_format set to JSON and validate the output against the expected schema before any user sees it. Validation checks: confirm every claim in cited_claims has a non-empty source_chunk_id that matches a provided chunk; confirm abstained_claims is populated when overall_abstention is true; reject outputs where confidence is high but source_chunk_id is missing. Retry logic: if validation fails, re-invoke the prompt once with the validation errors appended to [CONSTRAINTS]. If the second attempt also fails, route to a human review queue with the original query, retrieved chunks, and both failed outputs attached.

Model choice and latency: Use a model with strong instruction-following and JSON mode support (GPT-4o, Claude 3.5 Sonnet, or equivalent). This prompt is not suitable for small or fast models that tend to hallucinate citations under pressure. Expect 1-3 seconds of latency in a typical RAG pipeline. Logging and audit: Log every invocation with the full prompt, retrieved chunks, raw model output, validation result, and final routed decision. For high-risk domains, store these logs immutably for audit review. Human review integration: When overall_abstention is true or validation fails, surface the query and retrieved chunks in a review queue with the model's abstention reasoning pre-filled. Let human reviewers confirm the abstention or provide a manual answer. Never auto-retry into silence—every abstention event should be observable and actionable.