Answer Regeneration with Mandatory Citations Prompt

This prompt is designed for production RAG pipelines where an initial answer has been generated without adequate source grounding. Its primary job is to take that uncited answer, along with the original retrieved evidence, and regenerate a new answer where every factual claim is anchored to an inline citation. The ideal user is an AI engineer or backend developer integrating a cited-answer system into a product where user trust, auditability, or compliance depends on traceable claims. You need the original user query, the uncited model answer, and the set of retrieved source passages as inputs.

Use this prompt when your validation layer has detected missing citations, when a citation coverage score falls below a defined threshold, or when a human reviewer flags an answer as unsupported. It is not a general-purpose answer generator; it assumes the retrieval step has already occurred and that the evidence is available. Do not use this prompt for initial answer generation, for real-time conversational flows where latency is critical and a full regeneration pass is too slow, or when the retrieved evidence itself is known to be irrelevant or toxic—fix retrieval quality first. The prompt enforces a strict constraint: claims that cannot be sourced must be removed, not rephrased to sound more plausible.

Before wiring this into your application, define your citation completeness scoring rubric and your minimum acceptable threshold. The prompt includes a scoring instruction, but your harness should independently validate the output by checking that every sentence containing a factual assertion has a corresponding citation marker. In high-stakes domains like legal, finance, or healthcare, always route regenerated answers through human review before they reach the end user. The next section provides the copy-ready template you can adapt to your evidence schema and citation format.

The following prompt template is designed to be dropped into a production RAG pipeline after an initial answer has been generated without citations or with incomplete sourcing. It forces the model to rebuild the answer from scratch using only the provided evidence, inserting citation markers inline and explicitly removing any claim that lacks a direct source. The template uses square-bracket placeholders that your application must populate before sending the request to the model.

textCopy
You are a citation-verification assistant operating inside a production RAG pipeline. Your task is to regenerate an answer with mandatory inline citations anchored to the provided evidence. You must not introduce any claim that cannot be directly traced to a source passage.

## INPUT

**Original Answer (may contain unsourced claims):**
[ORIGINAL_ANSWER]

**Retrieved Evidence Passages (each with a unique source ID):**
[EVIDENCE_PASSAGES]

**Citation Format Requirement:**
[CITATION_FORMAT]

**Output Schema:**
[OUTPUT_SCHEMA]

## CONSTRAINTS

1. Every factual claim in the regenerated answer must be immediately followed by an inline citation referencing a source ID from the evidence passages.
2. If a claim from the original answer cannot be supported by any evidence passage, remove it entirely. Do not rephrase, soften, or hedge it.
3. If the evidence is insufficient to answer the question, produce an abstention response explaining what evidence was sought, what was found, and why it was insufficient.
4. Do not fabricate source IDs. Only use IDs present in the provided evidence passages.
5. Do not paraphrase a source passage and present it as a direct quote. If quoting, use exact text from the evidence and wrap it in quotation marks.
6. If multiple evidence passages support the same claim, cite the strongest passage. If passages conflict, note the conflict explicitly and cite both.

## CITATION COMPLETENESS SCORING

After regenerating the answer, append a `citation_completeness` object with these fields:
- `total_claims`: number of discrete factual claims in the regenerated answer
- `cited_claims`: number of claims with at least one citation
- `uncited_claims`: number of claims without any citation (must be zero)
- `removed_claims`: number of claims from the original answer that were removed due to lack of evidence
- `conflicts_detected`: number of conflicting evidence pairs identified
- `abstention`: boolean indicating whether the answer is an abstention

## EXAMPLES

[FEW_SHOT_EXAMPLES]

## RISK LEVEL

[RISK_LEVEL]

To adapt this template for your pipeline, replace each placeholder with concrete values. [ORIGINAL_ANSWER] should contain the uncited answer your system generated. [EVIDENCE_PASSAGES] must be a structured list of source passages, each with a unique ID that your citation format can reference. [CITATION_FORMAT] defines the inline citation style—for example, [Source: {id}] or a numeric superscript. [OUTPUT_SCHEMA] should specify the expected JSON structure if your pipeline requires machine-readable output alongside the cited text. [FEW_SHOT_EXAMPLES] should include at least two worked examples showing a correct regeneration and a correct abstention. [RISK_LEVEL] should be set to high, medium, or low to control how aggressively the model removes unsourced claims. For high-risk domains such as healthcare or legal, set [RISK_LEVEL] to high and add a human-review gate after regeneration.

Before deploying this prompt, validate it against a golden dataset of answers with known citation gaps. Measure whether the regenerated output preserves all and only the claims that have evidence support. Common failure modes include the model inserting plausible-looking but fabricated source IDs, retaining unsourced claims by softening their language, or over-citing every sentence even when a single citation covers a paragraph. If your pipeline handles regulated content, always route the regenerated output through a human reviewer and log the citation_completeness object for audit trails. Do not treat this prompt as a substitute for retrieval quality—if the evidence passages are irrelevant or incomplete, the model will correctly abstain, and your retrieval step needs improvement.

This prompt is designed to sit inside a citation recovery harness—a pipeline stage that activates when a primary answer generator produces output with missing, incomplete, or unverifiable citations. The harness should intercept the uncited answer and the original retrieved evidence set, then invoke this regeneration prompt as a corrective step before the answer reaches the user. Do not call this prompt on every request; reserve it for outputs that fail a citation completeness check, a source-to-claim alignment validator, or a hallucination detector. The harness must preserve the original user query, the full retrieval context, and any citation schema requirements so the regeneration prompt has everything it needs to rebuild a properly cited answer.

Wiring the pipeline: Structure the harness as a sequential recovery chain with explicit validation gates. First, run a citation completeness check (such as the Citation Completeness Check Retry Prompt) to detect missing or malformed citations. If the check fails, extract the uncited answer and the original evidence passages, then call this regeneration prompt with [UNCITED_ANSWER], [RETRIEVED_EVIDENCE], [CITATION_SCHEMA], and [CITATION_REQUIREMENTS] populated. The prompt will produce a regenerated answer with inline citations and a citation coverage map. Before returning the result, validate the output against the schema: every citation marker must resolve to a passage in the evidence set, every factual claim must have at least one citation, and no claim should be present that cannot be sourced. If validation fails again, increment a retry counter and either re-invoke with stricter constraints or escalate to a human review queue. Model choice: Use a model with strong instruction-following and long-context handling (GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro). Set temperature low (0.0–0.2) to minimize creative drift during regeneration. Enable structured output mode if your provider supports it, binding the response to a schema that includes regenerated_answer, citation_map, and removed_claims fields.

Retry budget and escalation: Define a hard retry limit—typically 2–3 attempts—before escalating. On the first retry, use the same evidence set but add explicit [CONSTRAINTS] that tighten citation requirements (e.g., 'every sentence containing a factual assertion must end with a citation marker'). On the second retry, expand the retrieval window using the Evidence Gap Identification and Fill Prompt to pull in additional context, then regenerate. If the third attempt still fails validation, route the original answer, the failed regenerations, and the evidence set to a human review queue with a structured escalation payload. Log every attempt with the prompt version, model, retry count, validation results, and final disposition. This trace data feeds your prompt observability dashboard and helps you tune retry budgets over time. What to avoid: Do not silently drop uncited claims—always either cite them or explicitly list them as removed. Do not allow the regeneration prompt to introduce new factual content not present in the evidence. And do not skip the post-regeneration validation gate; an unvalidated regenerated answer is just as risky as the original uncited output.