Evidence Ranking with Diversity Constraint Prompt

This prompt is designed for RAG system designers and search engineers who need to rank retrieved evidence passages by relevance while enforcing diversity across sources, viewpoints, or document types. Standard relevance ranking often produces top-K homogenization where the highest-scoring passages all originate from a single authoritative source, repeating the same perspective and missing critical context from other documents. This prompt solves that problem by making diversity an explicit ranking constraint alongside relevance. Use it when your downstream answer generation or citation selection requires broad coverage, when you need to surface dissenting or complementary viewpoints, or when audit requirements demand multi-source corroboration.

The ideal user has already retrieved a candidate set of passages—typically 10 to 50—from a vector store, keyword index, or hybrid retrieval pipeline. They need to reduce this set to a ranked top-K (often 5 to 10 passages) that will fit within a context window for answer generation. The prompt expects you to provide the original query, the candidate passages with source identifiers, and a diversity dimension specification (e.g., source domain, author, publication, viewpoint). You should also configure the diversity strength parameter: a low setting permits minor source repetition when relevance is overwhelmingly strong, while a high setting forces strict source separation even at the cost of slightly lower relevance scores. The output is a ranked list with relevance scores, diversity annotations, and a coverage summary showing which sources and perspectives are represented.

Do not use this prompt when you need pure relevance ranking without diversity constraints, when your retrieval set is already small and naturally diverse (fewer than 5 passages from distinct sources), or when latency budgets cannot accommodate the additional reasoning tokens this prompt requires. The diversity constraint adds approximately 30-50% more output tokens compared to a standard ranking prompt because the model must explain trade-offs between relevance and source variety. For real-time applications with sub-second latency requirements, consider running diversity-aware ranking as an offline or async preprocessing step rather than inline during user request handling. Also avoid this prompt when your retrieval pipeline already enforces diversity at the retrieval stage through source-aware sampling or federated search across collections—in those cases, a standard relevance ranking prompt is sufficient.

This prompt template is designed for RAG system builders who need ranked evidence that covers multiple perspectives rather than converging on a single source. It accepts a query, a set of candidate passages with metadata, and configurable diversity constraints. The output is a ranked list where each tier balances relevance against source variety, preventing the common failure mode where all top-K passages originate from the same document or author. Use this when your downstream answer generation or citation selection requires broad coverage, such as in research synthesis, competitive analysis, or decision-support systems where one-sided evidence produces misleading outputs.

textCopy
You are an evidence ranking system that prioritizes both relevance and diversity.

## INPUT
Query: [QUERY]
Candidate Passages:
[PASSAGES]
  - Each passage includes: id, text, source_id, source_type, author, publication_date

## DIVERSITY CONSTRAINTS
- Minimum distinct sources required in top-K: [MIN_SOURCES]
- Maximum passages per source in top-K: [MAX_PER_SOURCE]
- Diversity dimensions to enforce: [DIVERSITY_DIMENSIONS]
  (Options: source_id, source_type, author, viewpoint, document_section)
- Total passages to return (K): [TOP_K]

## RANKING CRITERIA
1. Relevance: How directly the passage addresses the query, including specificity and factual alignment.
2. Authority: Source credibility based on publication recency, author expertise, and document type.
3. Diversity contribution: Whether the passage adds coverage from an underrepresented source or perspective.

## OUTPUT SCHEMA
Return a JSON object with this exact structure:
{
  "ranked_passages": [
    {
      "rank": 1,
      "passage_id": "string",
      "relevance_score": 0.0-1.0,
      "authority_score": 0.0-1.0,
      "diversity_contribution": "high|medium|low",
      "selection_rationale": "One sentence explaining why this passage was selected at this rank."
    }
  ],
  "diversity_summary": {
    "sources_represented": ["source_id"],
    "coverage_gaps": ["dimensions with insufficient representation"],
    "constraint_violations": ["any constraints that could not be satisfied"]
  },
  "excluded_passages": [
    {
      "passage_id": "string",
      "exclusion_reason": "redundant|low_relevance|constraint_conflict",
      "note": "Brief explanation"
    }
  ]
}

## CONSTRAINTS
- If fewer than [MIN_SOURCES] distinct sources have relevant passages, include all available and flag the gap in coverage_gaps.
- Do not fabricate diversity. If the candidate set lacks diverse perspectives, report it honestly.
- When relevance scores are tied, prefer the passage that improves source diversity.
- Exclude passages that are near-duplicates of higher-ranked selections.

## INSTRUCTIONS
1. Score every passage for relevance and authority independently.
2. Build the ranked list iteratively: select the highest-scoring passage, then select the next passage that maximizes both relevance and diversity contribution.
3. Stop when K passages are selected or no more passages satisfy the constraints.
4. Populate the diversity_summary and excluded_passages fields with honest assessments.

To adapt this template for your system, replace each square-bracket placeholder with concrete values from your retrieval pipeline. The [PASSAGES] block should be populated with your actual retrieval results, including the metadata fields your system tracks. If you don't have source_type or author metadata, remove those from the diversity dimensions options and adjust the input schema accordingly. For production deployments, validate the output JSON against the schema before passing ranked passages to downstream answer generation. When operating in high-stakes domains such as medical or legal evidence ranking, add a human review step that inspects the diversity_summary for coverage gaps before the ranked list is used. The most common adaptation mistake is setting [MIN_SOURCES] too high for narrow queries where only one or two sources exist—this forces the model to include irrelevant passages to satisfy an impossible constraint. Start with [MIN_SOURCES]=2 and [MAX_PER_SOURCE]=3 for a [TOP_K] of 5, then tune based on your retrieval corpus characteristics.