Hallucination Prevention via Mandatory Citation Prompt

This prompt is designed for production Retrieval-Augmented Generation (RAG) systems where unsupported claims are unacceptable. Its core job is to enforce a strict contract: every factual assertion the model generates must be explicitly tied to a retrieved source passage, and the model must refuse to produce statements when evidence is missing. The ideal user is an engineering lead or AI builder integrating a question-answering feature into a regulated, legal, financial, or clinical product where outputs must be auditable and verifiable. It assumes you have already retrieved a set of candidate passages and need the model to synthesize an answer that is fully grounded in those passages, not in its pre-training data.

Use this prompt when your application requires traceable, evidence-backed answers and you have a retrieval pipeline in place. It is appropriate for workflows like contract analysis, financial report Q&A, clinical guideline lookup, and compliance document review. The prompt explicitly instructs the model to abstain from answering when the provided context is insufficient, making it suitable for high-risk domains where a wrong answer carries more cost than no answer. You should pair this prompt with a post-generation validation harness that checks for citation presence, verifies that cited claims are actually supported by the referenced passages, and routes outputs with missing or fabricated citations for human review.

Do not use this prompt for creative writing, brainstorming, or tasks where source retrieval is not part of the architecture. It is not suitable for general chatbot conversations, open-ended ideation, or scenarios where the model is expected to draw on its internal knowledge. Avoid using it when the retrieved context is known to be noisy, irrelevant, or incomplete without a corresponding abstention strategy, as the prompt will cause the model to refuse to answer frequently. For lower-risk applications where traceability is optional, consider a lighter-weight citation prompt that encourages but does not mandate source grounding.

This prompt template is designed to be placed in the system message of a RAG pipeline where unsupported claims are unacceptable. It instructs the model to treat every factual assertion as requiring a direct citation to a retrieved source passage. The prompt uses square-bracket placeholders that you replace at runtime: [CONTEXT] for the retrieved documents, [CITATION_FORMAT] for your required citation style, [OUTPUT_SCHEMA] for structured output requirements, and [RISK_LEVEL] to adjust the strictness of the abstention behavior. The template is self-contained and can be pasted directly into your system prompt field without modification beyond variable substitution.

textCopy
You are an evidence-grounded assistant operating in a high-trust domain. Your responses must be verifiable against provided source material only.

## CORE RULE: MANDATORY CITATION
Every factual claim, statement of fact, number, date, statistic, or assertion you make MUST be immediately followed by a citation to the specific source passage that supports it. You are forbidden from generating any factual statement that you cannot cite from the provided context.

## CITATION FORMAT
Use the following citation format for every claim: [CITATION_FORMAT]

## ABSTENTION RULE
If the provided context does not contain sufficient evidence to support a claim, you MUST NOT make that claim. Instead, explicitly state: "I cannot confirm this from the provided sources." Do not guess, infer beyond the text, or draw on external knowledge.

## OUTPUT STRUCTURE
Your response must follow this structure:
1. Answer the user's question using ONLY information from the provided context.
2. Every sentence containing a factual claim must end with a citation marker.
3. If multiple sources support a claim, cite all supporting sources.
4. If the context contains conflicting information, present both sides with their respective citations and note the conflict.
5. End your response with a "Citation Coverage Summary" listing which claims were cited and flagging any questions you could not answer from the sources.

## PROVIDED CONTEXT
Below is the only source material you may use. Do not reference, imply, or draw upon any information outside this context.

[CONTEXT]

## OUTPUT SCHEMA
Your final output must conform to this schema:
[OUTPUT_SCHEMA]

## RISK LEVEL
Current risk level: [RISK_LEVEL]
- If RISK_LEVEL is "high": You must cite every sentence, including transitional statements that contain implied facts.
- If RISK_LEVEL is "medium": You must cite every substantive factual claim; stylistic or transitional sentences may omit citations if they contain no factual content.
- If RISK_LEVEL is "low": You must cite key factual claims; minor background context may be stated without citation if it is directly observable from the provided context.

To adapt this template for your application, replace the square-bracket placeholders with your runtime values. For [CONTEXT], inject your retrieved passages with unique identifiers so citations can be traced back to specific chunks. For [CITATION_FORMAT], specify your required format such as [Source: {id}] or ({id}, §{section}). For [OUTPUT_SCHEMA], provide a JSON schema or structured format description if you need machine-readable outputs; otherwise, replace this section with your plain-text formatting rules. For [RISK_LEVEL], set the value based on your domain's tolerance for uncited claims—use "high" for regulated industries like healthcare, legal, or finance where every statement must be traceable. Before deploying, test this prompt against a golden dataset of questions with known answerable and unanswerable cases to verify that the model refuses to generate claims when evidence is missing and that all cited claims actually appear in the referenced passages.

This prompt is not a standalone artifact; it is a contract that must be enforced by the application layer. The prompt instructs the model to refuse uncited claims, but the harness is responsible for verifying that the model actually complied. In production, you will wrap this prompt in a pipeline that validates the output, retries on failure, logs every attempt, and escalates to a human when the model cannot produce a fully cited answer. The harness is what turns a well-designed prompt into a reliable, auditable system.

Start by defining a strict output schema that the model must follow. A JSON structure with a claims array works well: each claim object should contain a statement string, a citations array of source IDs, and a confidence enum. Before the model sees the prompt, inject the retrieved sources with unique IDs into the [CONTEXT] placeholder. After generation, run a citation validator that checks: (1) every claim has at least one citation, (2) every citation ID exists in the provided context, and (3) no claim is marked as uncited unless it is an explicit abstention. If validation fails, feed the error details back into a retry prompt: "The previous response contained claims without citations. Revise the answer so every factual statement cites a source ID from the provided context." Limit retries to a configurable maximum (typically 2-3) to avoid infinite loops.

For high-risk domains, add a human review trigger when the model abstains or when citation completeness falls below a threshold. Log the full prompt, the retrieved context, the raw model output, and the validation result to an audit store. This traceability is essential for regulated environments where you must prove that an answer was grounded in evidence. Choose a model that reliably follows structured output instructions—GPT-4o and Claude 3.5 Sonnet perform well here—and set temperature to 0 to minimize creative drift. Finally, build an eval harness that runs a golden set of questions with known source coverage and measures citation precision, citation recall, and hallucination rate. Run this eval on every prompt or context assembly change before deploying.