Conversational Citation Persistence Prompt Template

This prompt is for RAG system builders who need source attribution to remain stable across multiple conversation turns. The core job is generating an answer that uses the exact same citation format, source identifiers, and reference style established in prior messages. This matters when users are cross-referencing answers, when audit trails require consistent source tracking, or when a UI component expects a deterministic citation schema that must not drift between turn one and turn ten.

Use this prompt when your application has already defined a citation contract—such as bracketed numeric references like [1], source-key identifiers like [doc_alpha], or inline URLs—and you need the model to persist that format without reinterpreting it. The prompt is designed for multi-turn RAG copilots, chat assistants, and support tools where a user might ask a follow-up question and expect the new answer's citations to be directly comparable to the previous answer's citations. It is not a citation format designer; it assumes the format already exists in conversation history.

Do not use this prompt when you are defining citation format for the first time, when you need to extract and normalize citations from unstructured text, or when the conversation history contains conflicting citation styles that require resolution. This prompt also does not handle source deduplication across turns or decide which sources to cite—it focuses strictly on format persistence. For high-stakes domains where citation errors could cause audit failures, always pair this prompt with a post-generation validation step that checks citation format consistency against prior turns before the answer reaches the user.

This prompt template enforces citation format stability and source reference continuity across a multi-turn RAG conversation. It is designed for systems where a user may ask follow-up questions and the assistant must cite sources using the exact same identifiers and format established in prior turns. The template uses square-bracket placeholders that you replace with your application's specific inputs, output schema, constraints, and examples before sending the request to the model. The core mechanism is a dedicated [CITATION_REGISTRY] that tracks previously cited sources so the model can reuse them consistently.

textCopy
SYSTEM:
You are a research assistant answering questions from a knowledge base. You must cite sources using the exact format and identifiers established in prior turns.

CITATION RULES:
- Use inline citations in the format [Source: DOC_ID] immediately after each claim.
- If a source was cited in a prior turn, reuse the exact same DOC_ID. Do not create a new identifier for the same document.
- If a new source is introduced, assign it a new DOC_ID using the same naming convention as prior turns.
- If a prior answer cited a source that is also relevant to this turn, reference it by its existing DOC_ID.
- End every answer with a "Sources" section listing each cited DOC_ID with its full title and a brief relevance note.

[CITATION_REGISTRY]:
[PRIOR_CITATIONS]

[CONVERSATION_HISTORY]:
[HISTORY]

[RETRIEVED_CONTEXT]:
[CONTEXT]

[USER_QUESTION]:
[QUESTION]

[OUTPUT_SCHEMA]:
{
  "answer": "string (the answer with inline citations)",
  "sources": [{"doc_id": "string", "title": "string", "relevance": "string"}],
  "new_citations": ["string (list of DOC_IDs introduced in this turn)"],
  "reused_citations": ["string (list of DOC_IDs reused from prior turns)"],
  "citation_consistency_check": "string (self-check confirming all DOC_IDs match the registry or follow the same convention)"
}

[CONSTRAINTS]:
- Do not fabricate sources. Only cite documents present in [RETRIEVED_CONTEXT].
- If no retrieved context supports an answer, respond with "I cannot find evidence to answer this question" and do not cite sources.
- Maintain the exact DOC_ID format used in [CITATION_REGISTRY]. If prior citations used "DOC-001", use "DOC-002" for the next new source, not "source_2" or "doc2".
- If the user's question is a follow-up, explicitly reference the prior answer's sources when they remain relevant.

To adapt this template, replace each square-bracket placeholder with your application's runtime data. [PRIOR_CITATIONS] should contain a structured list of all DOC_IDs and their titles from previous turns. [HISTORY] should include the full conversation transcript. [CONTEXT] should contain the retrieved passages for the current turn, each with a stable DOC_ID already assigned by your retrieval pipeline. [QUESTION] is the user's current message. The [OUTPUT_SCHEMA] can be simplified if you only need the answer text, but keeping the citation_consistency_check field provides a built-in self-audit that you can log or use to trigger a retry. For high-stakes domains, add a [RISK_LEVEL] placeholder that gates whether the model should abstain or escalate when citation consistency cannot be verified.

Integrating citation persistence into a multi-turn RAG system requires more than a single prompt call. The application layer must maintain a session-scoped citation registry that tracks source identifiers, document metadata, and the citation format used in prior turns. When a new user message arrives, the harness retrieves the active conversation history, extracts previously used citation formats and source IDs, and injects them into the prompt's [PRIOR_CITATION_FORMAT] and [PRIOR_SOURCE_IDS] placeholders. This registry should be stored server-side—not trusted to the client—and keyed by session ID. If the session is new or no prior citations exist, the harness passes an explicit empty state marker (e.g., `

NO_PRIOR_CITATIONS"`) so the model knows to establish a fresh format rather than hallucinate continuity."

After the model returns a response, the harness must validate citation persistence before the answer reaches the user. Implement a post-generation validator that checks: (1) every citation in the current response uses a format consistent with prior turns (exact match on bracket style, numbering scheme, or inline pattern), (2) source IDs that appeared in prior turns retain the same identifier (no renumbering or renaming), and (3) new sources introduced in this turn receive identifiers that don't collide with existing ones. If validation fails, trigger a retry with the validation error message appended to the prompt as [CORRECTION_FEEDBACK]. Limit retries to two attempts; if the third output still fails, log the failure, surface the best-effort answer with a warning flag, and alert the operations channel. For high-stakes domains, route failed validations to a human review queue instead of auto-escalating.

Model choice matters here. Smaller or older models often struggle with format consistency across turns, especially when the conversation exceeds 8-10 exchanges. Prefer models with strong instruction-following and long-context handling (e.g., Claude 3.5 Sonnet, GPT-4o, or fine-tuned variants). If you must use a smaller model, reduce the citation format complexity—stick to simple numeric references like [1] rather than compound formats like [DocA §3.2]. Log every turn's citation registry state, validation results, and retry count for observability. When debugging citation drift, trace backward through the session to find the exact turn where format divergence began. The most common production failure mode is not the prompt itself but the application layer silently dropping prior citation state during context window truncation—ensure your context assembly logic preserves the citation registry even when older dialogue turns are compressed or evicted.