Source Attribution with Inline Citation Prompt

This prompt is for RAG application developers who need every factual statement in a generated answer to carry a precise, verifiable citation back to a source chunk. It produces answers with inline citation markers such as [Doc3-P4] or [Chunk-12] that map directly to provided source material. Use this when provenance is non-negotiable: compliance reports, legal research, clinical summaries, or any product where users must verify where each claim came from.

This is not a general Q&A prompt. It assumes you have already retrieved relevant chunks and numbered them. The prompt enforces citation discipline, detects unsupported statements, and refuses to generate claims without evidence. Do not use this prompt for creative writing, open-ended brainstorming, or scenarios where the model should synthesize freely without source grounding. It is also a poor fit when your retrieval pipeline is unreliable—the prompt will expose gaps by refusing to answer rather than hallucinating, which is correct behavior but may frustrate users expecting fluent responses.

Before deploying this prompt, ensure your retrieval system produces consistently numbered chunks with stable identifiers. The citation format must survive chunk reordering, deduplication, and context-window truncation. If your application requires partial-match handling or evidence sufficiency scoring, pair this prompt with the Evidence Sufficiency Assessment Prompt or Evidence Matching with Partial Match Handling Prompt from this pillar. For agent systems that synthesize across multiple tool calls, use the Source Attribution for Agent-Generated Answers Prompt instead, as it handles attribution preservation across multi-step trajectories.

This template is the core instruction set for a Source Attribution with Inline Citation system. It instructs the model to synthesize an answer exclusively from provided context chunks and to attach a bracketed citation tag to every factual statement. The placeholders let you inject your retrieved documents, specify the citation format your downstream parser expects, and define the output structure. Copy the block below into your system prompt or user message, replace each square-bracket variable with your application's data, and wire it into your retrieval pipeline.

textCopy
You are a precise research assistant. Your task is to answer the user's question using ONLY the provided context chunks. You must cite the source of every factual statement you make.

## CONTEXT CHUNKS
[CONTEXT]

## CITATION FORMAT
For every factual claim, insert a citation tag immediately after the relevant sentence or clause. Use the format: [CITATION_FORMAT].
Example: "The Acme revenue grew 14% year-over-year [doc3_para2]. This was driven by new product lines [doc3_para4]."

## CONSTRAINTS
- Do not use any knowledge outside the provided context chunks.
- If the context does not contain the answer, state "The provided documents do not contain sufficient information to answer this question." Do not guess.
- If multiple chunks support the same claim, cite the most specific or authoritative one.
- If chunks contradict each other, state the contradiction explicitly and cite both sources.
- Do not fabricate citations. Every tag must correspond to a real chunk ID in the provided context.

## USER QUESTION
[INPUT]

## REQUIRED OUTPUT FORMAT
[OUTPUT_SCHEMA]

## ADDITIONAL INSTRUCTIONS
[CONSTRAINTS]

To adapt this template, start by populating [CONTEXT] with your retrieved chunks, each prefixed with a unique, stable identifier like [doc3_para2] or [chunk_17]. These IDs must survive your retrieval and chunking pipeline so the citation tags remain traceable. Set [CITATION_FORMAT] to the exact pattern your post-processing parser expects, such as [source_id] or (source_id). The [OUTPUT_SCHEMA] field should describe the expected JSON or markdown structure, for example: Return a JSON object with keys 'answer' (string with inline citations) and 'cited_sources' (array of unique source IDs used). Use [CONSTRAINTS] to add domain-specific rules, such as "Prefer peer-reviewed sources over blog posts" or "Flag any claim with only a single source." After pasting, test the prompt with a golden dataset of questions and known-good citations to validate that the model respects the format and does not hallucinate source IDs.

This prompt is designed to be the final generation step in a RAG pipeline, sitting directly after retrieval and re-ranking. It expects pre-fetched document chunks with stable identifiers (paragraph IDs, chunk indices, or section labels) already attached. Do not pass raw, unlabeled text and expect the model to invent stable citations. The application layer must prepare the context by injecting identifiers into each passage before assembly. The prompt's [CITATION_FORMAT] placeholder should be set to a machine-parseable pattern such as [source:{id}] or <cite id="{id}"/> so that a post-processing validator can extract and verify every citation against the provided source set.

Wire the prompt into a structured generation harness that enforces a strict output contract. Configure your model call to request a JSON response with a response_text field containing the answer with inline citations and a citations array listing every source_id used. After generation, run a citation validator that checks: (1) every citation tag in response_text corresponds to a source_id present in the input context, (2) no source_id in the citations array is hallucinated, and (3) every claim in the answer has at least one citation. If validation fails, route to a retry recovery prompt that includes the specific validation error message and the original context. Limit retries to two attempts before escalating to a human review queue with the full trace attached.

For model choice, prefer models with strong instruction-following and structured output support (GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro). Set temperature to 0 or near-zero to minimize citation format drift. Log every generation with the prompt template version, retrieved context IDs, generated citations, and validator results. This trace is essential for debugging hallucinated citations and for audit workflows that require provenance evidence. Avoid using this prompt in streaming mode without a citation accumulator that can validate citations after the full response is received—partial citation validation mid-stream will produce false negatives.