Passage Diversity Enforcement Prompt for Coverage

This prompt is designed for RAG pipeline engineers and search architects who observe that standard relevance ranking returns a monotonous set of top passages. In open-ended or multi-faceted queries—such as 'impact of climate change on agriculture'—a pure vector similarity search might return ten passages about drought stress in corn while ignoring economic, soil health, or pest-related dimensions. The job-to-be-done is not just retrieving relevant text; it is ensuring the final context window provides comprehensive, non-redundant evidence for downstream answer synthesis. The ideal user is someone who has already tuned their retrieval but sees diminishing returns in answer quality because the model is generating from a narrow evidence base.

You should use this prompt when your knowledge base spans broad domains, user queries are exploratory or comparative, or when downstream tasks require balanced evidence. It is particularly effective as a post-retrieval, pre-generation step in a RAG pipeline. However, do not use this prompt for factoid-style queries with a single correct answer (e.g., 'What is the boiling point of water?'), as diversity enforcement will artificially exclude highly relevant, corroborating passages. Avoid it in high-throughput, low-latency systems where the additional LLM call for diversity selection violates your SLOs; in those cases, a deterministic algorithm like Maximal Marginal Relevance (MMR) is a better fit. Also, do not use this prompt if your retrieval set is already small (fewer than 5 passages), as diversity selection requires a sufficient pool to choose from.

Before implementing, verify that your retrieval step is returning a sufficiently large candidate set (e.g., top 20-50 passages). If you are already using a re-ranker, insert this diversity prompt after re-ranking but before context assembly. The output of this prompt is a structured selection of diverse passages with explicit coverage justification, which you should log for offline evaluation. Start by running this against a golden dataset of queries with known facet coverage to calibrate the diversity threshold before enabling it in production.

This prompt template enforces passage diversity by instructing the model to select a set of retrieved passages that maximizes topic coverage rather than simply taking the top-scoring results. It is designed for knowledge-base RAG systems where users ask questions that span multiple subtopics, and where returning five near-identical passages about the same facet produces a narrow, unhelpful answer. The template requires a list of candidate passages with relevance scores, a target selection count, and a coverage justification for each chosen passage. Use this when your retrieval pipeline returns clustered results and you need to ensure the final context window represents the full breadth of the query's information needs.

codeCopy
You are a passage diversity enforcer for a RAG retrieval pipeline. Your task is to select a diverse set of passages from the candidate list that maximizes topic coverage for the given query.

## INPUT
- Query: [QUERY]
- Candidate Passages: [CANDIDATE_PASSAGES]
- Target Selection Count: [TARGET_COUNT]
- Diversity Criteria: [DIVERSITY_CRITERIA]

## CANDIDATE PASSAGES FORMAT
Each passage is provided as a JSON object with the following fields:
- id: unique passage identifier
- text: passage content
- relevance_score: pre-computed relevance score (0.0 to 1.0)
- source: document or section source

## INSTRUCTIONS
1. Review all candidate passages and identify distinct topics, facets, or information dimensions present in the set.
2. Select exactly [TARGET_COUNT] passages that maximize coverage across these distinct topics.
3. Prioritize diversity of information over raw relevance score. A slightly less relevant passage that covers an unrepresented topic is preferred over a highly relevant passage that duplicates already-covered content.
4. For each selected passage, provide a coverage justification explaining what unique topic or information dimension it contributes.
5. If fewer than [TARGET_COUNT] distinct topics exist in the candidate set, select only the available distinct passages and flag the coverage gap.

## DIVERSITY CRITERIA
[DIVERSITY_CRITERIA]

## OUTPUT SCHEMA
Return a JSON object with the following structure:
{
  "selected_passages": [
    {
      "id": "string",
      "coverage_justification": "string explaining what unique topic this passage covers",
      "relevance_score": number
    }
  ],
  "coverage_summary": {
    "topics_covered": ["list of distinct topics identified"],
    "topics_missing": ["list of query-relevant topics not covered by any candidate passage"],
    "diversity_score": number between 0.0 and 1.0 indicating how well the selection covers the query's information needs
  },
  "selection_rationale": "brief explanation of trade-offs made between relevance and diversity"
}

## CONSTRAINTS
- Do not select passages that are near-duplicates or cover the same information dimension.
- If two passages cover the same topic, select the one with the higher relevance score.
- Flag any information gaps where the candidate set lacks coverage for a query-relevant topic.
- Maintain source diversity where possible; avoid over-selecting from a single document.

Adapt this template by adjusting the [DIVERSITY_CRITERIA] placeholder to match your domain. For technical documentation, criteria might include "procedural steps, configuration options, troubleshooting guidance, and version compatibility." For legal research, criteria might include "statutory interpretation, case precedent, jurisdictional analysis, and procedural requirements." The [TARGET_COUNT] should align with your context window budget—typically 3 to 8 passages for most RAG pipelines. If your retrieval system does not provide pre-computed relevance scores, remove that field from the candidate format and adjust the selection logic to rely solely on diversity. Always validate the output against your expected schema before passing selected passages to the answer generation step. For high-stakes domains, route the coverage summary to a human reviewer when the diversity score falls below 0.5 or when critical topics appear in the missing list.

The Passage Diversity Enforcement Prompt is designed to sit between your retrieval step and your answer generation step in a RAG pipeline. It consumes a set of top-k retrieved passages and produces a diverse subset with coverage justifications. In production, this prompt is not a one-shot call—it requires a harness that validates the output structure, enforces topic coverage constraints, and provides fallback behavior when the model fails to produce a valid diverse selection. The harness should treat this prompt as a filtering and re-ranking stage with a strict output contract, not as an optional enrichment step.

Wire the prompt into your application as a post-retrieval processor. After your vector or hybrid search returns the top N passages (typically 20–50), pass them to the LLM with the diversity prompt template. The model should return a JSON object containing a selected_passages array (each with passage_id, summary, and topic_label fields) and a coverage_justification object mapping required topics to selected passages. Validation is critical: before passing the diverse subset to your answer generation step, validate that (1) the output is valid JSON, (2) every passage_id references an actual input passage, (3) the number of selected passages does not exceed your context budget, and (4) all required topics from your [REQUIRED_TOPICS] list are covered. If validation fails, implement a retry loop with a maximum of 2 attempts, appending the validation error to the retry prompt. After 2 failures, fall back to a deterministic diversity algorithm (e.g., Maximal Marginal Relevance) and log the incident for review.

For observability, log the diversity prompt's input passage count, output passage count, topic coverage ratio, and validation status on every call. Track the coverage gap rate—the percentage of calls where the model fails to cover all required topics—as a key health metric. If this rate exceeds 5%, investigate whether your retrieval step is failing to surface passages for certain topics or whether the prompt's topic definitions need refinement. For high-stakes applications where missing a topic could cause harm (e.g., medical or legal RAG), route coverage gaps to a human review queue rather than silently proceeding with incomplete context. Model choice matters: use a model with strong instruction-following and JSON output reliability (GPT-4o, Claude 3.5 Sonnet, or equivalent). Avoid smaller or older models that may collapse diverse selections into redundant top-relevance picks, defeating the purpose of this prompt.