Multi-Passage Nuance Preservation Prompt

This prompt is designed for RAG application developers and knowledge-base teams who need to generate answers from multiple source passages that contain subtle but critical differences in scope, certainty, or framing. The core job-to-be-done is producing a synthesized response that faithfully preserves qualifiers like 'preliminary evidence suggests,' 'in patients under 65,' or 'applies only to enterprise-tier accounts' rather than flattening them into a generic, overconfident summary. The ideal user is an engineer or AI builder who has already retrieved a set of relevant passages and needs the model to respect the boundaries and hedges present in each source, not just extract the most common keywords.

Use this prompt when your sources disagree on degree rather than fact—for example, when one study reports a 15% improvement with a specific caveat and another reports a 12% improvement in a different population. Do not use this prompt when sources directly contradict each other on core facts; that scenario requires a dedicated conflicting evidence resolution prompt. This prompt also assumes your retrieval pipeline has already filtered for relevance and that you are not asking the model to judge source authority from scratch. The required inputs are a set of pre-retrieved passages and a user query. Optional inputs include a risk level, domain-specific terminology constraints, and an output schema for structured nuance preservation.

Before deploying, validate that your retrieval pipeline is not introducing its own bias by over-representing one perspective. If the prompt will be used in regulated domains such as healthcare or finance, you must add a human review step and log the source passages alongside the generated answer for auditability. The next step is to copy the prompt template, adapt the placeholders to your data model, and run it against a golden dataset of passages with known nuance requirements to establish a baseline for eval checks.

This template is designed for RAG application developers and knowledge-base teams who need to synthesize answers from multiple retrieved passages while preserving the qualifiers, hedges, scope boundaries, and domain-specific distinctions present in the original sources. The prompt instructs the model to avoid flattening subtle differences into a generic summary and instead surface where sources agree, disagree, or express different levels of certainty. Use this template as a starting point—adapt the placeholders, add domain-specific terminology constraints, and pair it with the eval checks described in the testing section of this playbook.

textCopy
You are an evidence synthesis assistant. Your task is to combine information from multiple source passages into a single coherent answer while preserving the nuance, qualifiers, and distinctions present in each source.

## INPUT

**Question:** [USER_QUESTION]

**Source Passages:**
[SOURCE_PASSAGES]

## CONSTRAINTS

[CONSTRAINTS]

## INSTRUCTIONS

1. Read all source passages carefully. Identify where sources agree, where they disagree, and where they express different levels of certainty, scope, or framing.

2. Produce a synthesized answer that:
   - Preserves qualifiers such as "may," "likely," "in some cases," "according to," and domain-specific hedging language.
   - Distinguishes between claims supported by multiple sources and claims from a single source.
   - Explicitly notes when sources disagree or present conflicting information.
   - Does not fabricate consensus where none exists.
   - Cites the source passage ID for each factual claim using the format [Source: ID].

3. If a source passage contains information that is irrelevant to the question, omit it from the synthesis.

4. If the source passages are insufficient to answer the question fully, state what is known and explicitly identify what remains unknown or uncertain.

5. Do not introduce information not present in the source passages.

## OUTPUT FORMAT

[OUTPUT_SCHEMA]

## EXAMPLES

[EXAMPLES]

## RISK LEVEL

[RISK_LEVEL]

To adapt this template, replace each square-bracket placeholder with concrete values. For [SOURCE_PASSAGES], include each retrieved passage with a unique ID, its content, and optionally metadata such as recency or authority scores. For [CONSTRAINTS], add domain-specific rules such as terminology requirements, forbidden phrases, or citation formatting standards. For [OUTPUT_SCHEMA], define the expected structure—this could be a JSON schema with fields for the answer, confidence level, source mapping, and disagreement flags, or a simpler markdown format with inline citations. For [EXAMPLES], provide at least one few-shot example showing how nuance is preserved when sources differ in certainty or scope. For [RISK_LEVEL], specify whether the domain requires human review before the output reaches users. If you are using pre-scored passages, add a weighting instruction to the constraints section so the model prioritizes higher-authority or more recent sources. Always pair this prompt with a post-generation faithfulness verification step to catch unsupported claims before delivery.

This prompt is designed to sit between your retrieval engine and your end-user response surface. It expects a set of pre-ranked passages, a user query, and an output schema that enforces nuance preservation. In a typical RAG pipeline, you'll call this prompt after retrieval and relevance filtering but before final answer delivery. The prompt's primary job is to prevent the model from flattening qualifiers, hedges, and domain-specific distinctions into a generic summary. This means your implementation harness must validate that the output actually preserves those nuances, not just that it produces valid JSON.

Start by wrapping the prompt in a function that accepts query, passages (an ordered list of objects with text, source_id, and optional authority_score), and output_schema as a JSON Schema definition. Before calling the model, assemble the prompt with these inputs substituted into the [INPUT], [CONTEXT], and [OUTPUT_SCHEMA] placeholders. Choose a model with strong instruction-following and long-context handling—GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro are good starting points. Set temperature low (0.0–0.2) to reduce creative drift from the source passages. After receiving the response, run three validation passes: (1) structural validation against your output schema using a JSON Schema validator, (2) nuance-preservation validation using a secondary LLM call that checks whether qualifiers, hedges, and source-specific framing from the input passages appear in the output, and (3) faithfulness validation that verifies no claim in the output contradicts or extends beyond the input passages. If any validation fails, retry once with the validation errors appended to the prompt as [CONSTRAINTS] feedback. After two failures, escalate to a human review queue with the original passages, the failed outputs, and the validation error details attached.

Log every call with a trace ID that connects the retrieval set, the prompt version, the model response, and all validation results. This trace is essential for debugging nuance loss in production—without it, you won't know whether the model dropped a hedge, the retriever missed a key passage, or the validator was too strict. For high-stakes domains like healthcare or legal, add a human approval step before the response reaches the user, even when all validations pass. The approval interface should display the source passages side-by-side with the synthesized answer and highlight any hedges or qualifiers that the nuance-preservation validator flagged as borderline. Finally, avoid wiring this prompt into a streaming response path unless you've built a post-stream validation step that can block or amend the output before the user sees the full answer. Nuance preservation requires seeing the complete synthesis before you can verify it.