Passage Salience Ranking Prompt for Long Documents

This prompt is for document intelligence engineers and RAG pipeline architects who need to rank sections, chunks, or passages from a long document by their relevance to a specific query. It produces ordered salience scores with key excerpt identification and includes coverage checks to prevent missing critical sections. Use this when your retrieval step returns a large set of candidate passages from a single long document and you need to prioritize which sections to include in a context window for downstream answer generation or summarization.

This prompt is not a replacement for initial vector or keyword retrieval. It assumes you already have a candidate set of passages from a single document and need to rank them by salience before context assembly. The ideal user has already run a broad retrieval step—likely returning 20 to 50 candidate chunks from a contract, technical manual, regulatory filing, or research paper—and now faces a context window budget that can only hold 5 to 10 of those chunks. The prompt's job is to prevent the most relevant passage from being crowded out by superficially similar but less salient neighbors. It works best when the document has clear structural boundaries (sections, clauses, chapters) and the query requires synthesizing information that may be distributed across non-adjacent parts of the document.

Do not use this prompt when you need to compare passages across multiple documents—that requires cross-document evidence ranking with authority weighting, not single-document salience. Do not use it as a standalone retrieval step on a raw document without prior chunking and candidate retrieval. Do not use it when the query can be answered from a single obvious passage; the overhead of ranking is wasted when retrieval precision is already perfect. Do not use it for real-time streaming applications where the ranking latency would violate response time SLAs. Finally, do not use it without a coverage check: the prompt includes explicit instructions to flag any document sections that appear relevant but were missing from the candidate set, and skipping that check risks silent omissions in high-stakes document review workflows.

This prompt template is designed for document intelligence engineers who need to rank sections or chunks from long documents by their relevance to a specific query. It produces ordered salience scores with key excerpt identification and includes coverage checks to prevent missing critical sections. The template uses square-bracket placeholders that must be replaced with actual values before sending to the model. Each placeholder serves a specific purpose: [QUERY] defines what you're looking for, [PASSAGES] contains the document chunks to rank, [OUTPUT_SCHEMA] specifies the expected JSON structure, [CONSTRAINTS] enforces ranking rules and coverage requirements, and [RISK_LEVEL] adjusts the model's behavior for high-stakes domains.

textCopy
You are a document intelligence system that ranks passages from long documents by their relevance to a query. Your task is to produce an ordered salience ranking with supporting evidence and coverage verification.

## Query
[QUERY]

## Passages to Rank
[PASSAGES]

## Output Schema
Return a JSON object with this exact structure:
{
  "ranked_passages": [
    {
      "passage_id": "string",
      "salience_score": number (0.0 to 1.0),
      "rank": integer,
      "key_excerpt": "string (most relevant quote from the passage)",
      "relevance_rationale": "string (why this passage matters to the query)"
    }
  ],
  "coverage_assessment": {
    "query_aspects_covered": ["string"],
    "query_aspects_missing": ["string"],
    "overall_coverage_score": number (0.0 to 1.0),
    "recommendation": "string (whether additional retrieval is needed)"
  }
}

## Constraints
[CONSTRAINTS]

## Risk Level
[RISK_LEVEL]

## Instructions
1. Read all passages carefully before ranking.
2. Assign salience scores based on direct relevance to the query, not general topic similarity.
3. For each passage, extract the single most relevant quote as the key_excerpt.
4. Rank passages in descending order of salience (highest score = rank 1).
5. After ranking, assess whether all aspects of the query are covered by the ranked set.
6. Flag any query aspects that lack supporting passages.
7. If coverage is incomplete, recommend whether additional retrieval is needed.
8. Do not fabricate information not present in the passages.

To adapt this template for your use case, replace each placeholder with concrete values. For [QUERY], provide the exact question or search intent. For [PASSAGES], supply the document chunks as a structured list with unique IDs—this is critical for traceability in downstream systems. The [CONSTRAINTS] placeholder should be replaced with domain-specific ranking rules, such as prioritizing recent passages, penalizing redundant content, or enforcing minimum coverage thresholds. For [RISK_LEVEL], use 'low' for exploratory ranking, 'medium' for production RAG pipelines where ranking errors degrade answer quality, or 'high' for regulated domains where missing critical sections creates compliance risk. In high-risk scenarios, always add a human review step before the ranked output feeds into answer generation or decision support systems. Validate the output against your expected schema before passing it downstream, and log salience scores for observability into ranking quality over time.

Integrating the Passage Salience Ranking Prompt into an application requires more than a single API call. The prompt is designed to be a stateless scoring function within a larger document processing pipeline. The typical harness receives a long document, chunks it into sections or passages, and then calls the LLM with the prompt template for each chunk or in batches, depending on context window limits. The application layer is responsible for chunking strategy, deduplication, and assembling the final ranked list from individual LLM responses. Do not rely on the model to handle documents larger than its context window; pre-chunking is mandatory.

Validation and Output Contract: The prompt instructs the model to produce a JSON object with salience_score, key_excerpts, and rationale. The application harness must validate this schema strictly. Use a JSON schema validator to confirm the presence and types of all required fields. The salience_score must be a number between 0.0 and 1.0. If validation fails, implement a single retry by feeding the raw model output and the validation error back into a repair prompt (see the Output Repair and Validation Prompts pillar). After one retry, discard the chunk's score or flag it for human review rather than entering an infinite retry loop. Log all validation failures to monitor prompt drift over time.

Coverage Check and Assembly: A critical failure mode for long documents is missing a vital section entirely. After all chunks are scored, the harness should execute a coverage check. This can be a secondary LLM call that receives the original query, a list of all chunk titles or identifiers, and the top-N ranked chunks. The coverage check prompt asks: 'Are there any sections not in the top-N that are likely critical to answering the query?' If the coverage check flags a missed section, it should be inserted into the ranked list with an explanation. The final output of the harness is an ordered list of passages with scores, excerpts, and a coverage_verified boolean flag. For high-stakes document review, always route the final ranked list and coverage report to a human reviewer before using it to generate a downstream answer or summary.