Citation Rules to Grounded Answer Example Prompt

This prompt is for RAG system builders and AI engineers who need to convert written citation policies into concrete, few-shot examples that teach a model how to produce grounded, sourced answers. The job-to-be-done is translating abstract rules—such as 'cite every factual claim,' 'do not quote more than two sentences,' or 'flag conflicting sources'—into input-output demonstration pairs that a model can internalize without needing the full policy text in every request. The ideal user is someone who already has a retrieval pipeline and a citation policy document, and now needs to bake that policy into the model's behavior through examples rather than verbose system instructions.

Use this prompt when you have a stable citation policy and need to produce a set of question-to-answer demonstrations that exhibit correct citation behavior, including inline references, quote length limits, source conflict resolution, and answer abstention when evidence is insufficient. The prompt requires several inputs: the full citation policy text, a set of retrieved source passages with identifiers, and representative user questions. It produces structured demonstrations with the question, the grounded answer, and a faithfulness check confirming that every claim in the answer is supported by the provided sources. Do not use this prompt when your citation policy is still in flux, when you lack source material to ground the answers, or when you need real-time answer generation rather than static example creation.

Before running this prompt, ensure your citation policy is finalized and your source passages are clean and attributable. The output examples should be reviewed by a domain expert for faithfulness and policy compliance before being added to a production prompt library. After generating the examples, run them through an LLM judge or manual review to confirm that inline citations match source content, quote limits are respected, and conflict resolution follows the stated policy. Treat these examples as golden test cases for regression testing whenever the citation policy or retrieval pipeline changes.

This template transforms a set of citation rules and a question-answer pair into a demonstration example that teaches the model how to produce sourced, faithful answers. Use it when you need to build a few-shot example set that enforces inline citation formatting, quote length limits, source conflict resolution, and abstention behavior—without writing verbose procedural instructions. The prompt expects you to provide the raw source material, a question, and a target answer style so it can generate a complete demonstration pair.

textCopy
You are an expert at creating few-shot demonstrations for retrieval-augmented generation (RAG) systems. Your task is to convert a set of citation rules and a question-answer pair into a single, self-contained demonstration example that teaches a model how to produce grounded answers.

## INPUTS
- Citation Rules: [CITATION_RULES]
- Source Documents: [SOURCE_DOCUMENTS]
- Question: [QUESTION]
- Target Answer: [TARGET_ANSWER]
- Output Schema: [OUTPUT_SCHEMA]
- Constraints: [CONSTRAINTS]
- Risk Level: [RISK_LEVEL]

## INSTRUCTIONS
1. Read the Citation Rules carefully. Identify the required citation format, quote length limits, source attribution style, conflict resolution policy, and abstention criteria.
2. Review the Source Documents and the Question. Note which passages support the Target Answer and which passages contradict it or are irrelevant.
3. Construct a demonstration example consisting of:
   - A `user` message containing the question and the source documents (formatted as they would appear in a real RAG prompt).
   - An `assistant` message containing the answer that strictly follows every citation rule.
4. The assistant answer must:
   - Include inline citations in the format specified by the Citation Rules.
   - Limit direct quotes to the maximum length specified in the Citation Rules.
   - Explicitly resolve any conflicts between sources, citing the conflicting passages.
   - Abstain or express uncertainty if the Citation Rules require it for insufficient or conflicting evidence.
   - Match the structure defined in the Output Schema.
5. If the Risk Level is HIGH, include a `[NEEDS HUMAN REVIEW]` marker at the start of the assistant answer and flag any claims that cannot be fully verified against the provided sources.
6. Return ONLY the demonstration example in the following format:

## DEMONSTRATION EXAMPLE
**user:** [Formatted question and source documents]
**assistant:** [Grounded answer with inline citations]

After copying the template, replace each square-bracket placeholder with concrete values. For [CITATION_RULES], specify the exact format (e.g., [Source ID, Paragraph]), maximum quote length in words, and conflict resolution policy. For [SOURCE_DOCUMENTS], include the full text of each passage with its identifier. For [OUTPUT_SCHEMA], define the expected structure—such as Answer paragraph with inline citations, followed by a Sources section listing each cited document ID. For [RISK_LEVEL], set to HIGH for regulated domains like healthcare or legal, which triggers the human-review marker and forces explicit uncertainty language. Before deploying the generated demonstration, validate it against your citation policy: check that every claim in the assistant answer is traceable to a source passage, that no quote exceeds the length limit, and that conflicts are surfaced rather than silently resolved.

This prompt is designed to be a data generation step within a larger RAG system, not a real-time user-facing endpoint. Its job is to convert a static citation policy document into a reusable set of few-shot examples that teach a downstream answer-generation model how to cite sources correctly. The implementation harness must treat this prompt as a batch processor: it takes a policy, a set of source documents, and a target count, and it produces validated, reviewed example pairs. Do not call this prompt on every user query; instead, run it offline when the citation policy changes, when new source types are introduced, or when the downstream model's citation behavior drifts.

Wire the prompt into a generation-and-validation loop with the following stages. First, load the citation policy and a curated set of source documents that represent the diversity of your knowledge base—include short passages, long documents, conflicting sources, and sources with missing information. For each source, call the prompt to generate a candidate question-answer pair with inline citations. Immediately run a faithfulness check: use a separate LLM judge or a deterministic string-match validator to confirm that every cited claim in the answer appears verbatim or in close paraphrase within the cited source span. Flag any example where a citation points to a source that does not contain the claim. Discard or queue for human repair any example that fails this check. This stage is non-negotiable for production use; a hallucinated citation in a training example will teach the downstream model to fabricate sources.

After faithfulness validation, run a policy compliance check against the original citation rules. Verify that quote length limits are respected, that source conflict resolution follows the policy (e.g., preferring more recent sources or flagging contradictions), and that the model correctly abstains or expresses uncertainty when the source material is insufficient. Log every generated example with its validation results, the source document ID, and the policy version used. Store examples in a versioned example registry that maps to the specific citation policy revision. When the policy changes, re-run the generation pipeline and compare new examples against the previous set to detect behavioral drift before the examples reach the downstream prompt.

For high-stakes domains—legal, medical, financial—add a human review gate before any generated example enters the downstream prompt's few-shot set. Present reviewers with the source document, the generated question, the answer, and the faithfulness validation results side by side. Allow reviewers to approve, edit, or reject each example. Track reviewer decisions to identify systematic failure modes in the generation prompt. If more than 10% of examples require edits, tune the prompt or improve the source document selection before the next batch run. Finally, wire the approved example set into your downstream RAG prompt's few-shot block, ensuring the examples are ordered by difficulty and diversity to maximize teaching transfer. Monitor the downstream model's citation accuracy in production; if citation errors increase, trigger a re-generation and review cycle.