Citation Density Control System Prompt

This prompt is for product teams and AI engineers who have already built a RAG system that generates answers with citations, but are now facing a readability problem: the output is either a wall of bracketed source IDs that users ignore, or a smooth narrative with no clear evidence trail. The job-to-be-done is to programmatically control how many citations appear per paragraph or per claim, so the system balances trust and readability without manual post-editing. The ideal user is an engineering lead or product manager who can define a density target (e.g., 'at least one citation per paragraph, no more than three per sentence') and needs the model to follow that constraint consistently across varying answer lengths and topic complexities.

Use this prompt when you have a working citation format (inline brackets, footnotes, or structured objects) and need to add a density control layer on top. It is not a replacement for a citation formatting prompt or an evidence selection pipeline. Do not use this prompt if your retrieval quality is poor or your evidence ranking is unreliable—density control amplifies existing problems by forcing citations to weak sources. This prompt also assumes you can measure density programmatically after generation; if you cannot count citations per structural unit (paragraph, claim, sentence), you cannot enforce the constraint. The prompt works best with models that have strong instruction-following capabilities (GPT-4 class, Claude 3.5 Sonnet, or equivalent) and degrades on smaller models that struggle with quantitative output constraints.

Before deploying this prompt, define your density bounds explicitly: minimum citations per paragraph to prevent under-attribution, maximum citations per sentence to prevent clutter, and a policy for what counts as a citation (inline reference, footnote marker, structured object). Pair this prompt with a post-generation validator that counts citations against your structural units and flags violations. In high-stakes domains like healthcare or legal, add a human review step for any output that violates density constraints, and log density metrics over time to detect drift. The next section provides the copy-ready template you can adapt to your specific density targets and output format.

This prompt template enforces a configurable citation density policy for RAG answer generation. Instead of treating citation count as an afterthought, it bakes density targets directly into the model's instruction contract. The system prompt defines a target range (e.g., 1–3 citations per paragraph), instructs the model to self-measure density before emitting output, and triggers corrective behavior when density falls outside bounds. Use this when your product team needs to balance readability with auditability—too few citations erode trust, too many clutter the answer.

textCopy
You are an answer-generation system that cites source documents with inline references.

[CITATION_FORMAT_RULES]

[DENSITY_POLICY]
- Target citation density: [MIN_CITATIONS] to [MAX_CITATIONS] citations per paragraph.
- A paragraph is any block of text separated by one or more blank lines.
- Count each unique source reference exactly once per paragraph, even if cited multiple times.

[MEASUREMENT_PROTOCOL]
Before finalizing your answer:
1. Count the number of paragraphs in your draft.
2. Count the number of citations in each paragraph.
3. Flag any paragraph with fewer than [MIN_CITATIONS] citations as UNDER-CITED.
4. Flag any paragraph with more than [MAX_CITATIONS] citations as OVER-CITED.

[CORRECTION_RULES]
- For UNDER-CITED paragraphs: add citations from [SOURCE_DOCUMENTS] that support the claims made. If no supporting source exists, either remove the unsupported claim or mark it with [UNCERTAINTY_MARKER].
- For OVER-CITED paragraphs: retain only the [MAX_CITATIONS] most relevant or authoritative citations. Prefer citations with higher [RETRIEVAL_SCORE] or more recent [DOCUMENT_DATE].

[OUTPUT_REQUIREMENTS]
- Produce the corrected answer with inline citations in [CITATION_STYLE] format.
- Append a density report section:

[DENSITY_REPORT] Paragraphs: [COUNT] Under-cited paragraphs corrected: [COUNT] Over-cited paragraphs corrected: [COUNT] Final density range: [MIN]–[MAX] citations/paragraph

codeCopy
- If any claim remains uncited after correction, include an [UNCITED_CLAIMS] section listing them.

[SOURCE_DOCUMENTS]
[CONTEXT]

[INPUT]

Adapt this template by tuning the density range to your domain. Legal and clinical answers may require tighter lower bounds (2–4 citations per paragraph), while customer-facing FAQ responses might permit 0–1 with an explicit common-knowledge exemption. Replace [CITATION_FORMAT_RULES] with your inline format spec—bracketed IDs, footnote markers, or structured JSON citation objects. The [UNCERTAINTY_MARKER] placeholder should resolve to a phrase like "[Source needed]" or "[Unverified]" that your downstream UI can render distinctly. Wire the density report into your observability pipeline to track over-citation and under-citation rates over time, and set alerts when the correction count spikes, which often signals retrieval degradation or policy misalignment.

Integrating the Citation Density Control prompt into a production RAG system requires more than dropping it into the system message. The prompt works best as a pre-generation policy layer that sits between retrieval and answer synthesis. When the retriever returns a set of candidate passages, the application should first calculate the expected citation density based on the number of claims the retrieved evidence can support. This pre-computed density target—expressed as a range like '1–2 citations per paragraph' or 'at least one citation per factual claim'—is then injected into the [DENSITY_TARGET] placeholder before the prompt is assembled. Without this application-side calculation, the model receives only vague density instructions and produces inconsistent results across different input lengths.

Validation and retry logic is essential because citation density failures are measurable. After the model generates an answer, a post-processing validator should count the number of citations against the number of paragraphs and factual claims. If the density falls below the configured minimum threshold, the application should trigger a retry with an explicit density-failure message appended to the prompt, such as: The previous answer had only 2 citations across 5 paragraphs. Minimum required: 1 citation per paragraph. Regenerate with proper attribution. For over-citation—where every sentence has a bracketed source and readability collapses—the validator should check whether citation-to-sentence ratio exceeds a maximum threshold (e.g., 0.7) and retry with a readability-priority instruction. Implement these checks as deterministic post-processing functions, not as additional LLM calls, to keep latency and cost predictable.

Model choice and tool integration affect density control behavior. Models with strong instruction-following (GPT-4, Claude 3.5 Sonnet) reliably enforce density constraints when the target is explicit and numeric. Smaller or older models may drift toward either under-citation or excessive bracketed references. For RAG systems using tool-calling architectures, consider exposing citation density as a configurable parameter in the tool schema rather than embedding it only in the system prompt—this allows per-request density tuning without prompt rewrites. When the system serves multiple use cases (e.g., legal review requiring dense citations vs. customer support preferring sparse attribution), store density profiles as named configurations (legal_strict, support_light) and select them at runtime based on the request context. Log every answer's citation count, paragraph count, density ratio, and any retry events to your observability platform so the team can track density drift and adjust thresholds without deploying new prompts.