Multi-Source Evidence Synthesis Prompt Template

This prompt template is designed for RAG application developers who need to combine multiple retrieved passages into a single, coherent answer. It is the generation step in a standard RAG pipeline, positioned after retrieval and re-ranking but before output validation or human review. The primary job-to-be-done is faithful synthesis: producing a unified response with inline citations, preserved nuance across sources, and explicit handling of source agreement or conflict. Use this when your retrieval pipeline returns 3–15 relevant passages and you need a cohesive answer rather than a disjointed patchwork of excerpts.

The ideal user is an AI engineer or product developer building a citation-aware application over a knowledge base. Required context includes the user's question and the set of retrieved passages, each with a unique source identifier. The prompt assumes your retrieval step has already returned relevant passages. If your retrieval quality is poor or your sources are untrusted, fix those upstream before applying this synthesis prompt. This template is not a replacement for retrieval quality—it amplifies good retrieval and exposes bad retrieval through unfaithful or contradictory outputs.

Do not use this prompt when you have a single source passage (a simpler QA prompt suffices), when the retrieved passages are irrelevant or adversarial, or when the user expects a creative or opinion-based response rather than an evidence-grounded answer. This prompt is also inappropriate for real-time safety-critical decisions without human review. For high-stakes domains such as healthcare, legal, or finance, always pair this synthesis step with a downstream faithfulness verification prompt and a human-in-the-loop approval gate before any action is taken on the output.

This template is designed to be pasted directly into your system or user prompt for a large language model. It instructs the model to act as an evidence analyst, synthesizing multiple retrieved passages into one unified answer. The core instruction set forces the model to preserve nuance, explicitly handle agreement and disagreement between sources, and provide inline citations for every claim. Before using this template, ensure you have already retrieved and ranked your evidence passages; this prompt handles synthesis, not retrieval. The template uses square-bracket placeholders that you must replace with your application's specific values before sending the request.

textCopy
You are an evidence synthesis analyst. Your task is to combine the provided evidence passages into a single, coherent, and faithful answer to the user's question. Follow these rules strictly:

1.  **Ground Every Claim:** Every factual statement in your answer must be directly supported by at least one of the provided evidence passages. Do not introduce outside knowledge.
2.  **Preserve Nuance:** Maintain qualifiers, hedges, and domain-specific language from the sources. Do not flatten subtle differences into a generic summary.
3.  **Handle Conflicts Explicitly:** If sources disagree on a point, do not fabricate consensus. Clearly state the disagreement, present the conflicting evidence from each side, and cite the sources for each position.
4.  **Use Inline Citations:** After every sentence or clause derived from a source, include a citation tag like [Source: [SOURCE_ID]]. Use the exact Source ID provided in the context.
5.  **Identify Gaps:** If the provided evidence is insufficient to fully answer the question, state what is missing and provide only a partial answer based on what is known.
6.  **Output Format:** Return your response as a valid JSON object matching the schema below.

---
## USER QUESTION
[USER_QUESTION]

## EVIDENCE PASSAGES
[EVIDENCE_PASSAGES]

## OUTPUT SCHEMA
{
  "answer": "The synthesized answer string with inline citation tags like [Source: ID].",
  "evidence_summary": "A brief summary of the key points from the evidence, noting areas of agreement and conflict.",
  "gaps_identified": ["A list of questions or information gaps the evidence could not address."],
  "confidence": "high | medium | low"
}

## CONSTRAINTS
- Do not invent facts or citations.
- If the evidence is contradictory, the 'confidence' field must be 'low' or 'medium'.
- The 'answer' field must contain citation tags.

To adapt this template, replace the [USER_QUESTION] and [EVIDENCE_PASSAGES] placeholders with your application's runtime data. The [EVIDENCE_PASSAGES] should be a formatted string that includes a unique [SOURCE_ID] for each passage, which the model will use for inline citations. You can extend the [CONSTRAINTS] section to add domain-specific rules, such as 'Prioritize sources with an authority score above 0.8' or 'If a source is older than one year, flag it as potentially stale.' For high-risk domains like healthcare or finance, add a constraint requiring the model to explicitly state that its output is not professional advice and must be reviewed by a human. The output schema is a strong contract; validate the JSON structure and the presence of citation tags in your application layer before showing the answer to a user. A common failure mode is the model omitting citations for claims it considers 'common knowledge'—the explicit instruction to ground every claim is the primary defense against this.

The Multi-Source Evidence Synthesis prompt is designed to sit at the final stage of a RAG pipeline, after retrieval, relevance filtering, and evidence ranking have already completed. It expects a curated set of ranked passages, not a raw retrieval dump. Before calling this prompt, your application should assemble the [PASSAGES] array with each passage's text, source identifier, and any pre-computed relevance or authority scores. The prompt then produces a unified answer with inline citations, preserved nuance, and explicit handling of source agreement or conflict. This is not a retrieval prompt—it assumes the evidence is already selected and ranked upstream.

Wire the prompt into your application as a post-retrieval synthesis step with a validation layer immediately after the model response. The validation layer should parse the output for required structural elements: inline citations that map to source identifiers, a conflict or agreement section when sources diverge, and a faithfulness check that verifies every claim in the answer traces back to at least one passage in [PASSAGES]. For high-stakes domains, implement a secondary LLM-as-judge evaluation using the Synthesis Faithfulness Verification prompt from the same pillar to score the output before it reaches the user. If the faithfulness score falls below a configurable threshold (typically 0.85 for production), route the response to a human review queue rather than delivering it directly. Log every synthesis call with the input passages, the generated answer, the faithfulness score, and any human override for audit and prompt improvement cycles.

For retry logic, implement a two-tier recovery strategy. If the model output fails structural validation—missing citations, unparseable format, or refusal to synthesize—retry once with the same prompt and a stronger [CONSTRAINTS] block that explicitly requires the missing elements. If the second attempt also fails, fall back to a simpler per-source summary rather than attempting synthesis. For faithfulness failures detected by the judge prompt, do not retry with the same passages; instead, flag the passage set for retrieval quality review and return a qualified partial answer with explicit gap flags. Model choice matters here: use a model with strong instruction-following and long-context handling, such as Claude 3.5 Sonnet or GPT-4o, because synthesis quality degrades sharply with weaker models that lose track of source boundaries. Avoid using this prompt with models below roughly 70B parameters unless you have validated synthesis faithfulness on your specific domain data.

When deploying in regulated or customer-facing contexts, never deliver a synthesized answer without running the faithfulness verification step. Log the full trace—retrieved passages, their scores, the prompt template version, the model response, the faithfulness verdict, and any human review decision—for audit and governance. If your use case involves conflicting sources, configure the application to surface conflicts rather than silently resolving them; the prompt includes conflict-handling instructions, but your UI should also present disagreement indicators to users. Finally, treat this prompt as versioned product code: maintain a changelog, run regression tests against a golden set of passage-answer pairs, and gate production updates behind eval pass rates on faithfulness, coverage, and hallucination metrics.