Comparative Multi-Source Analysis Prompt Template

This prompt is designed for decision-support RAG applications where a user asks about a single topic and the system has retrieved documents from multiple sources. Instead of synthesizing a single answer that might hide disagreement, this prompt instructs the model to produce a structured comparison that surfaces agreements, disagreements, and unique insights from each source. Use it when the goal is to inform a decision by making the evidence landscape visible, not to produce a single authoritative answer. It assumes you have already retrieved relevant passages from at least two distinct sources and that each source is identifiable with metadata such as title, author, or date.

This prompt belongs in the answer-generation stage of a RAG pipeline, after retrieval and before any final formatting for a UI. It is not a replacement for retrieval, evidence ranking, or fact-verification steps. Do not use this prompt when you need a single synthesized answer, when sources are not individually identifiable, or when the user expects a definitive recommendation rather than a balanced comparison. It is also inappropriate for real-time fact-checking where source credibility must be assessed algorithmically before comparison—this prompt assumes all provided sources are equally admissible and leaves credibility judgments to the reader.

Before wiring this prompt into production, ensure your retrieval step returns passages with stable source identifiers. The prompt's value collapses if the model cannot attribute claims to specific sources. Plan to pair this prompt with an evaluation rubric that checks for balanced representation, accurate attribution, and absence of hallucinated agreements or disagreements. In high-stakes domains such as legal or clinical decision support, always route the output through human review before it reaches the end user.

This template is the core instruction set for a comparative multi-source analysis. It is designed to be placed directly into the system or user message of a model call. The prompt forces the model to structure its output around agreements, disagreements, and unique insights, preventing the common failure mode of a bland summary that erases conflict. Before using this template, you must have already retrieved a set of documents or passages about a single topic from your knowledge base. The prompt assumes the retrieved context is provided in the [CONTEXT] placeholder, with each source clearly delimited and labeled for citation.

codeCopy
You are an analytical assistant that compares and contrasts information from multiple sources. Your task is to produce a structured comparison of the provided sources on the topic defined by the user.

## INPUT
[CONTEXT]

## USER QUERY
[QUERY]

## INSTRUCTIONS
1. **Identify the Core Topic:** Based on [QUERY] and [CONTEXT], define the central subject of the comparison in one sentence.
2. **Extract Key Points:** For each source in [CONTEXT], extract the key claims, data points, or arguments relevant to the topic.
3. **Structured Comparison:** Organize your analysis into the following sections:
   - **Points of Agreement:** List claims or facts where sources concur. Cite the agreeing sources for each point using their labels.
   - **Points of Disagreement:** List claims or facts where sources conflict. For each conflict, state the opposing positions and cite the sources holding each position. Do not resolve the conflict; present it neutrally.
   - **Unique Insights:** For each source, list any significant claim or piece of information not found in the other sources.
4. **Synthesis and Confidence:** Provide a brief synthesis of the overall picture, explicitly noting the level of consensus or conflict. State your confidence in the analysis, flagging any areas where evidence is thin or sources seem unreliable.

## CONSTRAINTS
- Do not add information not present in [CONTEXT].
- Cite sources using their exact labels from [CONTEXT] for every factual claim.
- If [CONTEXT] contains only one source, state that a comparison is not possible and provide a summary instead.
- If [CONTEXT] is empty or irrelevant to [QUERY], respond with "Insufficient context to perform the analysis."
- Maintain strict neutrality; do not favor one source over another.

## OUTPUT FORMAT
[OUTPUT_SCHEMA]

To adapt this template, replace the placeholders with your application's specific data. [CONTEXT] should be a string containing all retrieved documents, each prefixed with a unique label like [Source A], [Source B], etc. The [QUERY] is the user's original question. The [OUTPUT_SCHEMA] placeholder is critical for programmatic consumption; replace it with a strict schema definition, such as a JSON schema or a precise markdown structure, to enforce the sections defined in the instructions. For high-risk domains like legal or financial analysis, add a [RISK_LEVEL] constraint that mandates a human-readable disclaimer and requires a human review step before the output is considered final. The next step is to wire this prompt into a harness that validates the output against your schema and checks for citation integrity.

The Comparative Multi-Source Analysis prompt is designed to sit downstream of a retrieval step that fetches documents about a single topic from multiple sources. In a production harness, you should not pass raw, unsorted retrieval results directly into the prompt. Instead, pre-process the retrieved chunks by grouping them by source, deduplicating near-identical passages, and attaching stable source identifiers (e.g., [Source A], [Source B]) that the prompt can reference. The prompt expects a [TOPIC] and a [SOURCES] block containing the grouped, attributed evidence. If your retrieval pipeline returns more than 5-7 distinct sources, consider filtering to the most relevant or authoritative ones before invoking this prompt to keep the analysis focused and the token budget manageable.

Wire the prompt into your application with a structured output contract. The prompt template should include an [OUTPUT_SCHEMA] placeholder that specifies a JSON schema with fields like agreements (array of points where sources concur), disagreements (array of conflicting claims with source attribution), unique_insights (per-source novel information), and overall_assessment (a brief synthesis noting the balance of evidence). After receiving the model response, validate the JSON structure programmatically. Check that every claim in agreements and disagreements references at least one valid source identifier from the input. If validation fails, retry once with a repair prompt that includes the validation error message and the original context. For high-stakes decision-support applications, route outputs with a high disagreement-to-agreement ratio or low source coverage to a human review queue before surfacing to end users.

Model choice matters for this workflow. Use a model with strong instruction-following and long-context handling, such as Claude 3.5 Sonnet or GPT-4o, as the task requires holding multiple document perspectives in memory and comparing them systematically. Enable structured outputs or JSON mode if available to reduce parsing failures. Log every invocation with the topic, source count, validation status, and a hash of the output for auditability. Avoid using this prompt when the retrieved sources are all from the same author or organization, as the comparative framing may invent false disagreements to satisfy the instruction. In those cases, fall back to a single-source synthesis prompt instead.