Citation Density Control Prompt

This prompt is for product teams who have already built a working RAG pipeline and are now facing a specific production trade-off: their outputs are either so dense with citations that they are unreadable, or so fluent that no one can verify the claims. The job-to-be-done is to produce a single, configurable system instruction that lets the application layer control citation density at runtime, from minimal key-claim attribution to exhaustive per-sentence grounding, without rewriting the core prompt for every use case. The ideal user is an AI engineer or technical product manager who needs to ship a feature where the end-user's tolerance for citation noise varies by surface—think of a 'detail' toggle in a UI that actually changes model behavior.

Use this prompt when you have a single answer-generation flow that must serve multiple audiences or risk postures. For example, a customer-facing summary widget might need only one citation per paragraph, while an internal compliance review screen for the same content requires a citation on every date, figure, and proper noun. Do not use this prompt if your retrieval quality is poor or your source documents are unreliable; controlling citation density will not fix hallucinations, it will only change how they are presented. This prompt assumes that the underlying evidence is relevant and that your retrieval system can supply the passages needed to support dense citation when requested. If you are still debugging retrieval relevance or source authority, start with the Evidence Weighting and Scoring Prompt or the Source Gap Analysis Prompt instead.

Before integrating this prompt, define your density levels as explicit, testable contracts. A common three-tier schema is: 'minimal' (cite only quantitative claims and direct quotes), 'standard' (cite all factual claims that are not common knowledge within the domain), and 'exhaustive' (cite every declarative sentence). You will need to validate that the model actually changes its behavior across these levels, not just the number of brackets it produces. The next section provides the template, and the implementation harness section covers how to build a density validator that counts citation-to-sentence ratios and fails outputs that drift from the requested level.

This prompt template lets you configure how many citations appear in a generated answer, from minimal key-claim attribution to exhaustive per-sentence sourcing. Use it when your application needs to balance readability against traceability, and when different user roles or risk levels demand different citation densities. The template uses square-bracket placeholders that you replace at runtime with your retrieved context, user question, density configuration, and output schema.

textCopy
You are an evidence-grounded assistant. Your task is to answer the user's question using ONLY the provided retrieved context. You must cite your sources inline according to the specified citation density level.

## CITATION DENSITY LEVEL: [DENSITY_LEVEL]
- "minimal": Cite only the 1-3 most important factual claims. Do not cite common knowledge, definitions, or transitional statements.
- "moderate": Cite key claims, statistics, dates, names, and any claim a reader might reasonably question. Aim for 1 citation per 2-3 sentences on average.
- "exhaustive": Cite every factual statement, including definitions, background context, and supporting details. Every sentence that contains a fact from the sources must have a citation.
- "per-sentence": Attach at least one citation marker to every sentence, even if multiple sentences draw from the same source passage.

## CITATION FORMAT
Use numbered inline markers in square brackets: [1], [2], [3]. Each marker must correspond to a source in the provided context. If multiple sources support a claim, list all relevant markers: [1][3].

## RETRIEVED CONTEXT
[RETRIEVED_CONTEXT]

## USER QUESTION
[USER_QUESTION]

## OUTPUT REQUIREMENTS
1. Answer the question using only information from the retrieved context.
2. Apply the citation density specified above.
3. If the context is insufficient to answer, state that clearly and do not fabricate citations.
4. Return your answer in the following JSON format:
{
  "answer": "The full answer text with inline citation markers.",
  "citations": [
    {"marker": 1, "source_id": "doc_1", "passage": "The cited passage text."},
    {"marker": 2, "source_id": "doc_2", "passage": "The cited passage text."}
  ],
  "density_applied": "[DENSITY_LEVEL]",
  "citation_count": 3,
  "insufficient_context": false
}

## CONSTRAINTS
- Do not cite sources outside the retrieved context.
- Do not invent source passages or marker numbers.
- If [DENSITY_LEVEL] is "exhaustive" or "per-sentence", every factual sentence must have at least one citation marker.
- Preserve the exact passage text in the citations array.

To adapt this template, replace [DENSITY_LEVEL] with one of the four supported values based on your application's risk tolerance and user needs. Populate [RETRIEVED_CONTEXT] with your RAG pipeline output, including source IDs and passage text. Insert the [USER_QUESTION] at runtime. The JSON output schema is designed for programmatic consumption, making it straightforward to validate citation counts against the requested density level before surfacing the answer to users. For high-risk domains, add a human-review step when insufficient_context is true or when the actual citation_count falls below the expected threshold for the requested density.

The Citation Density Control Prompt is designed to be a configurable component in a production RAG pipeline, not a standalone chat instruction. The primary integration point is immediately after retrieval and before the final response is surfaced to the user. Your application should inject the retrieved context chunks, the user's query, and a runtime density configuration into the prompt template. The density configuration should be a controlled parameter—exposed as a user preference, a product feature flag, or an API parameter—with allowed values like minimal, balanced, and exhaustive. Never allow free-text density descriptions from untrusted input, as this can lead to prompt injection or unpredictable behavior.

After the model returns a response, you must validate the citation density before the output reaches the user. Implement a post-processing validator that counts the number of citation markers or reference brackets in the output and compares it against the expected density tier. For minimal density, expect citations only on key quantitative claims or direct quotes; for exhaustive density, expect a citation on nearly every sentence. If the validator detects a mismatch—such as zero citations in an exhaustive request or excessive citations that degrade readability in a minimal request—trigger a retry loop. The retry should resubmit the same prompt with an additional constraint injected into the [CONSTRAINTS] block, such as 'The previous output had insufficient citations. Every factual sentence must include a source reference.' Limit retries to a maximum of two attempts before falling back to a human review queue or surfacing the best-effort output with a warning flag.

Model choice significantly impacts this prompt's reliability. Models with strong instruction-following capabilities, such as Claude 3.5 Sonnet or GPT-4o, handle density instructions more consistently than smaller models. If you are using an open-weight model like Llama 3 or Mixtral, expect to invest more effort in few-shot examples within the [EXAMPLES] block to demonstrate each density tier. Log every invocation with the density configuration, the number of citations produced, the validator result, and the retry count. This telemetry is essential for tuning density thresholds and detecting drift when you change models or update the retrieval pipeline. For high-stakes domains like legal or healthcare, always route exhaustive density outputs to a human reviewer before they reach the end user, regardless of validator results.