Fact-Dense Passage Prioritization Prompt

This prompt is for retrieval-augmented generation (RAG) teams who are drowning in verbose, low-signal chunks. The core job-to-be-done is maximizing answer-relevant information per token before synthesis. You have a set of retrieved passages—often overlapping, padded with boilerplate, or heavy on rhetorical fluff—and you need to select the subset that carries the highest density of verifiable, query-relevant facts. The ideal user is an AI engineer or search engineer building a QA system over technical documentation, compliance manuals, or scientific literature where every token of context counts against a hard context-window budget. You already have a retrieval step that returns candidate passages; this prompt sits between retrieval and answer generation, acting as a fact-density filter.

Do not use this prompt when your primary concern is source credibility, temporal freshness, or multi-perspective balance. It is not a relevance scorer—it assumes passages are already topically relevant. It is not a deduplication tool, though it naturally penalizes redundant content. It is not appropriate for conversational RAG where turn history should influence selection, nor for comparative questions where balanced representation matters more than raw fact density. If your documents are primarily narrative, opinion, or marketing prose, fact density becomes a noisy signal; this prompt works best when passages contain discrete, extractable claims, measurements, parameters, or technical specifications. For legal or financial systems, pair this with a source-credibility prompt and a conflict-resolution step; fact density alone does not protect against authoritative-sounding falsehoods from a single low-quality source.

Before deploying, define what 'fact density' means for your domain. In API documentation, it might be parameter-count per sentence. In clinical text, it might be discrete findings per passage. In engineering specs, it might be measurable constraints per chunk. Write five examples of high-density and low-density passages from your own corpus and use them as few-shot demonstrations. The prompt's output should feed directly into a token-budgeted selection step: take the top-K passages until you hit your context limit, then route the selected set to your synthesis prompt. Always log the excluded passages and their density scores for debugging retrieval quality and prompt drift over time.

The following prompt template is designed to be dropped into a retrieval-augmented generation (RAG) pipeline after initial retrieval but before answer synthesis. Its job is to re-rank a set of candidate passages so that those with the highest concentration of unique, verifiable facts per token appear first. This maximizes the answer-relevant information available within a constrained context window. The template uses square-bracket placeholders for all dynamic inputs, making it straightforward to parameterize in code. Before using this prompt in production, you must define the expected output schema, validation rules, and evaluation criteria described in the sections that follow.

textCopy
You are a fact-density analyst. Your task is to re-rank a set of retrieved passages so that passages with the highest concentration of unique, verifiable facts appear first. Prioritize passages that introduce new information not already covered by higher-ranked passages. Penalize verbose, redundant, or filler-heavy content.

## INPUT

Query: [USER_QUERY]

Passages to rank:
[PASSAGE_LIST]

## DEFINITIONS

- **Fact**: A specific, verifiable claim that can be checked against the passage text. Examples include numeric values, dates, named entities, causal relationships, and technical specifications.
- **Fact Density**: The number of unique facts divided by the token count of the passage. Higher is better.
- **Information Gain**: The number of facts in a passage that are not already present in passages ranked above it.

## CONSTRAINTS

[CONSTRAINTS]

## OUTPUT FORMAT

Return a JSON object with the following structure:

[OUTPUT_SCHEMA]

## EXAMPLES

[EXAMPLES]

## INSTRUCTIONS

1. For each passage, count the distinct, verifiable facts it contains.
2. Compute a fact-density score: facts / token_count.
3. Sort passages by fact density in descending order.
4. When passages have similar density scores, break ties by prioritizing passages with higher information gain relative to already-ranked passages.
5. Exclude any passage that contains zero verifiable facts or is entirely redundant with higher-ranked passages. Mark excluded passages with an exclusion reason.
6. Return the ranked list with scores, justifications, and exclusion reasons in the specified output format.

To adapt this template for your system, replace each square-bracket placeholder with concrete values at runtime. [USER_QUERY] should be the original user question, used as a relevance anchor. [PASSAGE_LIST] should be a formatted list of retrieved passages, each with a unique identifier and its full text. [CONSTRAINTS] can include token budgets, minimum density thresholds, or domain-specific fact definitions. [OUTPUT_SCHEMA] must be a strict JSON schema your application can validate. [EXAMPLES] should include at least two few-shot demonstrations showing a passage list, the expected ranking, and the reasoning. If your domain involves regulated content, add a constraint requiring that all ranked facts be directly attributable to the source passage text, and log the full ranking trace for audit. Do not deploy this prompt without first running the evaluation harness described in the testing section of this playbook.

This prompt is designed to sit between retrieval and answer synthesis in a RAG pipeline. It accepts a set of retrieved passages and a user query, then re-ranks them by factual density—maximizing unique, verifiable claims per token while deprioritizing verbose or redundant content. The output is an ordered list of passage IDs with density scores and justifications. You should call this prompt after initial retrieval and before context window packing, using its output to select the top-K passages that fit your token budget.

Integration pattern: Wrap the prompt in a function that accepts query, passages (list of objects with id, text, and optional source_metadata), and top_k. The function should construct the prompt with these inputs, call your model, and parse the JSON output. Validation is critical: validate that every returned passage_id exists in the input set, that density scores are numeric and within expected ranges, and that justifications are non-empty strings. If validation fails, retry once with the validation errors appended to the prompt as additional context. After two failures, fall back to a simpler relevance-only ranking prompt or log the failure for manual review. Model choice: Use a model with strong instruction-following and JSON output capabilities (e.g., GPT-4o, Claude 3.5 Sonnet). Avoid smaller models that may conflate keyword frequency with factual density.

Observability and evals: Log every invocation with the input passage count, output ranking, density scores, and validation status. Run periodic eval checks comparing the prompt's density rankings against human judgments on a golden dataset of passage sets where factual density has been manually annotated. Watch for failure modes: the prompt may over-prioritize list-heavy passages, penalize necessary context-setting sentences, or miss implicit facts that require inference. If you observe these patterns, add counterexamples to the prompt's few-shot examples or adjust the density definition in the constraints. When not to use this prompt: Skip it when retrieval already returns fewer passages than your context window can hold, when latency budgets are extremely tight, or when the downstream task requires narrative coherence over information density. In those cases, a simpler relevance ranking or direct context packing is more appropriate.