Financial Document Materiality Ranking Prompt Template

This prompt is designed for audit automation and financial intelligence product teams that need to transform a flat set of retrieved passages from SEC filings, earnings call transcripts, or audit reports into a defensible, ranked evidence list. The core job-to-be-done is evidence triage: when a retrieval step returns dozens or hundreds of potentially relevant text chunks, an auditor or analyst cannot review them all with equal care. This prompt forces the model to prioritize passages by quantitative materiality, disclosure requirement severity, and temporal relevance, producing a ranked list where each item includes a specific, traceable justification grounded in the source text. The ideal user is an engineering lead or product manager building a financial copilot, audit workpaper generator, or disclosure analysis tool who needs a repeatable, testable ranking step before downstream summarization or human review.

Use this prompt when the cost of missing a material disclosure is high and you need a ranking that an auditor can defend. It is appropriate for workflows such as prioritizing risk factors in a 10-K before generating a risk summary, ranking MD&A statements by their impact on reported financial condition, or ordering footnote disclosures by the magnitude of the quantitative adjustment they describe. The prompt is not a replacement for a retrieval step; it assumes you have already extracted candidate passages using vector search, keyword retrieval, or a document parser. It is also not a final answer generator—its output is a ranked evidence set that feeds into a subsequent summarization, extraction, or review step. Do not use this prompt for real-time market sentiment analysis, general financial Q&A without source documents, or scenarios where the ranking criteria are purely qualitative and cannot be tied to specific text spans.

Before implementing this prompt, ensure your pipeline can provide the required inputs: the full text of each candidate passage, its source document identifier, and the document date. The prompt's value depends on the model's ability to identify and compare quantitative thresholds, disclosure categories, and dates within the provided text. If your retrieval step strips out numeric values, table data, or date references, the ranking quality will degrade significantly. In regulated audit contexts, the ranked output should be treated as a suggested prioritization, not a final determination; always include a human-review gate for high-severity disclosures, non-GAAP metrics, or passages where the model's justification references a threshold that requires professional judgment to interpret.

This template is designed to be placed directly into your model's system prompt or as a dedicated ranking instruction block. It expects a set of extracted financial passages and a query context, and it returns a ranked list with materiality scores, rationale, and flags for stale data or non-GAAP metrics. The square-bracket placeholders must be replaced with your specific data, schemas, and constraints before use.

textCopy
You are a financial document materiality ranking engine. Your task is to prioritize a set of extracted passages from financial documents (filings, earnings call transcripts, reports) based on their quantitative impact, disclosure requirement, and temporal relevance to the user's query.

## INPUT
- Query Context: [QUERY]
- Extracted Passages: [PASSAGES_LIST]
  Each passage includes: passage_id, source_document, source_date, extracted_text.
- Reporting Date: [REPORTING_DATE]
- Entity: [ENTITY_NAME]
- Risk Level: [RISK_LEVEL] (options: low, medium, high, critical)

## RANKING CRITERIA
Rank each passage on a scale of 1-10 for each dimension, then compute a weighted materiality score:
1. Quantitative Impact (weight 40%): Does the passage contain specific figures, percentages, or dollar amounts that materially affect financial position or performance?
2. Disclosure Requirement (weight 35%): Is this information required by GAAP/IFRS, SEC regulations, or exchange listing standards? Flag if the passage suggests a potential disclosure deficiency.
3. Temporal Relevance (weight 25%): How close is the source_date to the reporting_date? Is the information stale or superseded by more recent filings?

## OUTPUT SCHEMA
Return a JSON object with the following structure:
{
  "ranked_passages": [
    {
      "passage_id": "string",
      "materiality_score": number (1-10, weighted),
      "quantitative_impact_score": number (1-10),
      "disclosure_requirement_score": number (1-10),
      "temporal_relevance_score": number (1-10),
      "rationale": "string (concise explanation of ranking)",
      "flags": {
        "stale_data": boolean,
        "non_gaap_metric": boolean,
        "boilerplate_only": boolean,
        "requires_human_review": boolean
      }
    }
  ],
  "summary": {
    "total_passages_evaluated": number,
    "high_materiality_count": number (score >= 7),
    "stale_data_count": number,
    "non_gaap_flag_count": number,
    "overall_assessment": "string (brief summary of evidence quality and coverage)"
  }
}

## CONSTRAINTS
- Do not invent or assume financial figures not present in the extracted passages.
- If a passage contains only boilerplate language without entity-specific detail, set boilerplate_only to true and reduce the quantitative_impact_score accordingly.
- Flag any non-GAAP metric that is not clearly reconciled to a GAAP equivalent.
- If the source_date is more than 12 months before the reporting_date, set stale_data to true and reduce temporal_relevance_score.
- If risk_level is "high" or "critical", set requires_human_review to true for any passage with materiality_score >= 6.
- If no passages meet the materiality threshold, return an empty ranked_passages array and explain in overall_assessment.

To adapt this template, replace the placeholders with your actual data pipeline outputs. The [PASSAGES_LIST] should be a JSON array of passage objects from your retrieval or extraction step. The [QUERY] should capture the user's analytical intent, such as 'Identify material revenue recognition changes in Q4.' The [RISK_LEVEL] parameter can be set dynamically based on the user's role, the engagement type, or the entity's risk profile. After adaptation, run this prompt against a golden dataset of known materiality judgments to calibrate the weighting and flag behavior before production deployment.

This prompt is designed to operate as a post-retrieval ranking step inside a financial document intelligence pipeline. The typical flow is: (1) retrieve candidate passages from filings, earnings call transcripts, or internal reports using a hybrid search system, (2) pass the top-N passages through this prompt along with the query context and ranking criteria, and (3) consume the ranked, scored output downstream for answer generation, alerting, or audit workpaper assembly. The prompt expects a batch of passages and returns a materiality-ordered list with scores and reasoning. Do not use this prompt as a standalone Q&A system; it ranks evidence, it does not answer questions directly.

Integration pattern: Wrap the prompt in a service that accepts a PassageSet (list of objects with id, text, source_document, date) and a RankingContext (user query, materiality thresholds, reporting period). The service should: (1) validate that each passage has a non-empty text and a parseable date before calling the model, (2) call the model with a structured output schema (JSON mode or function-calling with a typed MaterialityRanking object), (3) validate the response schema and score ranges (0.0–1.0) on return, and (4) log the full prompt, response, and validation result for auditability. For high-stakes use cases such as SOX controls or earnings analysis, add a human review queue that surfaces passages where the model's materiality score exceeds a configurable threshold but the reasoning contains uncertainty markers like 'may', 'could', or 'unclear'.

Model choice and retries: Use a model with strong structured output support (GPT-4o, Claude 3.5 Sonnet, or equivalent) and enable JSON mode with a strict schema. Set temperature=0 for deterministic ranking. Implement a retry loop with up to 2 retries if the response fails schema validation or if the model returns a ranking_failure flag (indicating it could not parse the input). On persistent failure, log the raw passages and escalate to a human analyst rather than silently dropping the batch. For latency-sensitive pipelines, consider batching passages into groups of 10–15 and running parallel ranking calls, then merging and re-normalizing scores across batches using a simple min-max scaling step.

Eval and monitoring: Before shipping, build a golden dataset of 50+ passage sets with known materiality rankings (sourced from auditor annotations or SME review). Run the prompt against this dataset and measure Kendall's tau for rank correlation and precision@k for top-k material passages. Monitor in production for: (a) score distribution drift (sudden clustering at extremes), (b) stale data flag rate (passages older than the reporting period), and (c) non-GAAP metric flag rate (to catch when the model misses adjusted earnings or pro-forma figures). Wire these metrics into your observability stack and set alerts for deviations beyond 2 standard deviations from baseline. Avoid deploying this prompt on unvalidated retrieval output; garbage passages in produce garbage rankings out, regardless of prompt quality.