Inline Citation Marker Prompt for RAG Outputs

This prompt is for engineering teams building production RAG systems who need every factual claim in a generated answer to be traceable back to a specific retrieved passage. It produces answers with inline numbered citation markers and a corresponding reference list. Use it when your domain requires auditability, when downstream validators will check citation coverage, or when you need to separate direct quotes from paraphrased content. This prompt belongs after retrieval and before any post-generation citation validation step. It assumes you have already retrieved and formatted your source passages with unique identifiers.

Do not use this prompt when you lack a retrieval step, when your sources are not pre-segmented with stable IDs, or when the model must synthesize from its own training data. It is also the wrong choice for creative writing, open-ended brainstorming, or tasks where factual grounding is unnecessary. If your application needs structured JSON output with typed citation arrays rather than inline markers, prefer the Source-Aware Output Contract Prompt instead. For domains where wrong answers carry regulatory risk, pair this prompt with the Answer Abstention with Source Justification Prompt to handle insufficient-evidence cases gracefully.

Before deploying, implement a post-generation citation validator that checks every marker against the reference list and verifies that cited passages actually support the claims. Run the Citation Completeness Audit Prompt on a sample of outputs to catch fabrication risks. If your retrieval pipeline returns sources of varying quality, combine this prompt with the Evidence Weighting and Scoring Prompt to prioritize authoritative passages. Start with a moderate citation density and adjust based on whether your reviewers prioritize readability or full traceability.

The following prompt template is designed to be integrated directly into a production RAG pipeline. It instructs the model to produce a final answer where every factual claim is immediately followed by a bracketed citation marker (e.g., [1], [2]). These markers correspond to the index of the source passage in the provided [CONTEXT] block. The template enforces a strict separation between direct quotes and paraphrased summaries, and it includes a mandatory abstention rule for questions that cannot be answered from the provided sources. Use this template as the core instruction set for your generation step, adapting the [CONSTRAINTS] and [OUTPUT_SCHEMA] placeholders to fit your specific application's risk tolerance and data contract.

markdownCopy
You are an expert research assistant. Your task is to answer a user's question using ONLY the provided source passages. You must ground every factual statement with an inline citation marker.

## SOURCE PASSAGES
[CONTEXT]

## USER QUESTION
[INPUT]

## INSTRUCTIONS
1.  **Answer from Sources:** Generate a comprehensive answer based strictly on the information in the SOURCE PASSAGES. Do not use any outside knowledge.
2.  **Inline Citation Markers:** Insert a citation marker immediately after every factual claim, sentence, or data point derived from a source. Use the format `[Source Index]` where the index matches the number of the passage in the SOURCE PASSAGES list (e.g., `[1]`, `[2]`).
3.  **Quote vs. Paraphrase Tagging:**
    *   For direct quotes, use the marker `[Source Index, Q]` (e.g., `[3, Q]`).
    *   For paraphrased or summarized information, use the marker `[Source Index, P]` (e.g., `[3, P]`).
4.  **Abstention Rule:** If the provided SOURCE PASSAGES do not contain the information needed to answer the question, you must state clearly: "The provided sources do not contain sufficient information to answer this question." Do not guess or fabricate citations.
5.  **Output Format:** Structure your response as a single, coherent answer. Do not restate the instructions. The final output must be a valid JSON object matching the schema below.

## CONSTRAINTS
[CONSTRAINTS]

## OUTPUT SCHEMA
You must output a single JSON object with the following structure:
{
  "answer": "The full text of the answer with inline citation markers.",
  "citations": [
    {
      "marker": "[1, P]",
      "source_index": 1,
      "type": "paraphrase",
      "cited_text": "The exact passage text that supports the claim."
    }
  ],
  "abstained": false
}

To adapt this template, replace the [CONTEXT] placeholder with your retrieved and formatted passages, typically by iterating over your retrieval results and prepending an index number to each chunk. The [CONSTRAINTS] block is where you inject domain-specific rules, such as word limits, forbidden topics, or tone requirements. For high-risk applications, ensure your post-generation validation harness checks that every marker in the answer string has a corresponding entry in the citations array and that the abstained boolean is correctly set. The next step is to wire this prompt into an application harness that can validate its structured output and handle failure modes gracefully.

The inline citation marker prompt is not a standalone artifact—it is a component inside a larger RAG pipeline. The application layer is responsible for retrieving candidate passages, assembling the prompt with the correct [CONTEXT] and [QUERY] variables, and then validating the model's output before it reaches a user or downstream system. Treat the prompt as a function with typed inputs and a structured output contract. The harness around it should enforce that contract.

Start by defining the input schema. The [CONTEXT] placeholder expects a pre-ranked list of retrieved passages, each with a unique source identifier and the passage text. A typical structure is a JSON array of objects with id, source, text, and optional relevance_score fields, serialized into the prompt. The [QUERY] placeholder receives the user's original question. Before calling the model, validate that [CONTEXT] is not empty and that every passage has a non-null id—missing identifiers break citation mapping. If the retrieval step returns zero passages, bypass the model entirely and return a controlled abstention response rather than forcing the prompt to hallucinate citations.

After the model responds, run a post-generation validation step. Parse the output to extract all citation markers (e.g., [1], [2,3]) and verify that every marker references a passage ID that existed in the input context. Flag any orphaned citations for review. For high-trust domains, add a second validation pass that checks whether cited passages actually support the claims they are attached to—this can use a separate LLM judge prompt or a lightweight NLI model. Log every validation failure with the full prompt, raw output, and validation error for debugging. If citation completeness is critical, set a threshold (e.g., 90% of factual claims must have a citation) and route outputs below the threshold to a human review queue.

Model choice matters. This prompt works best with models that have strong instruction-following and long-context handling, such as Claude 3.5 Sonnet, GPT-4o, or Gemini 1.5 Pro. Smaller or older models may drop citation markers, merge passages, or ignore the marker placement rules. If you must use a smaller model, consider splitting the task: first generate the answer, then run a separate citation-attachment prompt that inserts markers into the completed text. For latency-sensitive applications, set a token budget on the output and use streaming with a post-processing step that buffers and validates the complete response before display.

Wire the prompt into your observability stack. Tag each request with a trace ID, the prompt version, the retrieval query, and the count of passages provided. Log the raw model output, the parsed citation list, and the results of every validation check. This trace data is essential for debugging citation failures, measuring hallucination rates over time, and running regression tests when you change the prompt or retrieval pipeline. If you are building a customer-facing product, never display uncited claims to users—always run the validation harness first, and fall back to a safe response if the output fails citation checks.