Evidence Ranking with Coverage Check Prompt

This prompt is designed for RAG system designers and search engineers who face a specific failure mode: a retrieval pipeline returns highly relevant passages, but they all address the same narrow aspect of a complex query. Standard ranking prompts optimize for relevance alone, which can produce a top-5 list where every passage discusses the same sub-topic while other required aspects of the query remain completely uncovered. The downstream answer generator then produces a response that appears well-grounded but is actually incomplete, missing critical information the user explicitly requested. This prompt solves that problem by performing two tasks simultaneously: ranking evidence by relevance and strength, and verifying that the ranked set collectively covers every required aspect of the user query. It produces a ranked list, a coverage map showing which query aspects are addressed, and explicit gap flags that trigger targeted re-retrieval before answer generation proceeds.

Use this prompt when answer completeness is a hard requirement and you cannot afford to generate responses from partial evidence. It is particularly valuable in domains where missing information carries high risk: medical question-answering where omitting a contraindication could cause harm, legal research where failing to address a jurisdictional split produces misleading advice, financial analysis where incomplete coverage of risk factors leads to poor decisions, and any customer-facing system where incomplete answers erode trust. The prompt is also useful during evaluation and QA workflows, where you need to audit whether your retrieval pipeline is surfacing sufficient coverage before blaming the generator for incomplete answers. However, do not use this prompt for simple factoid queries with a single correct answer, for latency-sensitive applications where the coverage check adds unacceptable overhead, or when your retrieval set is known to be comprehensive and coverage gaps are not a concern. In those cases, a standard ranking prompt with relevance scoring is more appropriate and cost-effective.

Before deploying this prompt, ensure you have a retrieval pipeline that can be re-queried with targeted sub-queries when gaps are detected. The prompt's gap flags are only useful if your system can act on them—either by issuing new retrieval calls for the uncovered aspects or by routing to a human reviewer with a clear explanation of what's missing. Also plan for the evaluation overhead: coverage verification requires defining what the required aspects of a query are, which may itself require an upstream decomposition step. If your queries are short and unambiguous, consider pairing this prompt with a query decomposition prompt that breaks the user's question into required sub-topics before ranking begins. Finally, treat the coverage map as a gating mechanism: if gaps are flagged, do not proceed to answer generation. Either re-retrieve, escalate to a human, or generate a response that explicitly acknowledges the coverage limitations rather than silently producing an incomplete answer.

This template provides a complete, self-contained prompt for ranking retrieved evidence passages by relevance and support strength while simultaneously checking whether the ranked set covers every aspect of the user's query. It is designed for RAG system builders who need to move beyond flat relevance ranking and into coverage-aware evidence selection. The prompt expects a query, a list of evidence passages with IDs, and an optional set of required sub-topics. If sub-topics are not provided, the model will extract them from the query before performing the coverage check. Use this template when your downstream answer generation depends on complete evidence coverage, and when missing information should trigger targeted re-retrieval rather than silent gaps.

Below is the copy-ready prompt template. Replace each square-bracket placeholder with your application's data before sending it to the model. The placeholders [QUERY], [EVIDENCE_PASSAGES], and [REQUIRED_SUBTOPICS] are required inputs. [CONSTRAINTS] and [OUTPUT_FORMAT] allow you to customize behavior and output shape without modifying the core ranking and coverage logic. For high-stakes domains such as healthcare, legal, or financial applications, set [RISK_LEVEL] to 'high' to enforce stricter coverage thresholds and require explicit uncertainty flags in the output.

textCopy
You are an evidence ranking and coverage analysis system. Your task is to rank the provided evidence passages by relevance and support strength for the given query, and then check whether the ranked set collectively covers all required sub-topics.

## INPUT

**Query:** [QUERY]

**Evidence Passages:**
[EVIDENCE_PASSAGES]

**Required Sub-Topics (if provided):** [REQUIRED_SUBTOPICS]

## CONSTRAINTS
[CONSTRAINTS]

## INSTRUCTIONS

1. If Required Sub-Topics are not provided, extract them from the Query by identifying all distinct aspects, entities, conditions, or questions that must be addressed for a complete answer.

2. For each evidence passage, assign a relevance score from 0.0 to 1.0 based on how directly it addresses the query and its sub-topics. Consider specificity, factual density, and source authority where indicated.

3. Rank all passages from highest to lowest relevance score. If multiple passages have identical scores, order them by specificity (more specific first).

4. For each required sub-topic, determine whether at least one passage in the ranked set provides sufficient evidence to address it. Mark each sub-topic as COVERED or GAP.

5. If any sub-topic has a GAP, flag it explicitly and recommend targeted re-retrieval queries that would fill the gap.

6. If [RISK_LEVEL] is 'high', apply a stricter coverage threshold: a sub-topic is only COVERED if at least two independent passages support it, or if a single passage provides explicit, unambiguous evidence with no hedging language.

## OUTPUT FORMAT
[OUTPUT_FORMAT]

## OUTPUT

Produce your response in the specified output format. Include the ranked evidence list with scores, the coverage map for each sub-topic, and any gap flags with recommended re-retrieval queries.

After pasting this template into your system, replace the placeholders with real data. For [EVIDENCE_PASSAGES], structure each passage as a JSON object or delimited text block containing an id, text, and optional metadata fields such as source, date, or authority_score. For [REQUIRED_SUBTOPICS], provide an array of strings representing the aspects that must be covered; if your application already performs query decomposition upstream, pass those decomposed sub-queries here. For [CONSTRAINTS], add any domain-specific rules such as 'prefer peer-reviewed sources over news articles' or 'penalize passages older than 2 years for time-sensitive queries.' For [OUTPUT_FORMAT], specify a JSON schema that includes ranked_passages (array of objects with id, score, rationale), coverage_map (object mapping each sub-topic to status and supporting_passage_ids), and gaps (array of objects with subtopic, recommended_query, and severity). Validate the model's output against this schema before passing ranked evidence downstream. For high-risk applications, add a human review step when any gap is flagged with 'high' severity, and log all coverage failures for audit and retrieval pipeline improvement.