Quote Freshness and Temporal Relevance Prompt

This prompt is for AI product engineers and RAG pipeline architects who need to prevent time-sensitive applications from serving stale evidence as fact. The job-to-be-done is assessing whether an extracted quote remains temporally valid for a query that carries an implicit or explicit time constraint. The ideal user is building a news summarization system, a financial intelligence dashboard, a regulatory monitoring feed, or any application where information decays in value over hours, days, or weeks. The reader should already have a retrieval pipeline that surfaces candidate quotes and source metadata including publication dates. This prompt sits between retrieval and answer generation, acting as a freshness gate before evidence is used downstream.

Use this prompt when your application handles queries where recency matters: earnings reports, breaking news, compliance updates, market prices, or operational incidents. The prompt expects a quote, its source publication timestamp, and the query's temporal sensitivity as inputs. It returns a structured freshness assessment with a score, a staleness warning flag, and reasoning that explains the temporal relationship. Do not use this prompt for atemporal knowledge (e.g., 'What is the capital of France?') or when source dates are unavailable. The prompt is designed for single-quote evaluation; for batch freshness scoring across multiple quotes, wrap it in an application-layer loop with aggregation logic.

Before deploying, test this prompt against known outdated evidence scenarios: a quote from last year for a query about today's stock price, a breaking-news quote from six hours ago when the situation has clearly evolved, or a regulatory quote from before a known rule change. The harness should verify that staleness warnings fire when expected and that freshness scores degrade appropriately as temporal distance increases. If your domain has specific freshness thresholds (e.g., financial data older than 15 minutes is stale, legal precedents remain valid for years), encode those as [CONSTRAINTS] in the prompt template rather than relying on the model's general world knowledge. For high-stakes domains like financial trading or clinical alerts, always route staleness-warned quotes to human review before they influence downstream decisions or user-facing answers.

This template is designed to be copied directly into your prompt management system or codebase. It accepts a user query, a set of candidate quotes with their source metadata, and a reference date, then outputs a structured freshness assessment for each quote. The prompt forces the model to reason about temporal relevance explicitly before scoring, which reduces false positives where a quote appears relevant but is actually stale.

textCopy
You are a temporal evidence analyst. Your job is to assess whether extracted quotes are fresh enough to support a time-sensitive query.

## INPUT
Query: [QUERY]
Reference Date (today): [REFERENCE_DATE]
Candidate Quotes:
[QUOTES]

Each quote includes:
- quote_id: unique identifier
- quote_text: the verbatim extracted text
- source_date: publication or document date in ISO 8601 format
- source_title: title of the source document
- source_authority: one of [high, medium, low, unknown]

## TASK
For each candidate quote, determine whether it is temporally valid for answering the query. Consider:
1. Does the query require recent information? Identify any time signals in the query (e.g., "latest," "current," "this quarter," specific date ranges).
2. Is the quote's source_date close enough to the reference date to be considered current for this query?
3. Could the information in the quote have changed since the source_date? Flag quotes where the subject matter is inherently volatile (e.g., stock prices, election results, product availability).
4. Does the quote contain temporal language that anchors it to a specific past period (e.g., "last year," "as of March," "previously")?

## OUTPUT SCHEMA
Return a JSON object with this exact structure:
{
  "query_time_sensitivity": "high|medium|low|none",
  "query_time_signals": ["list of time-related phrases found in the query"],
  "assessments": [
    {
      "quote_id": "string",
      "freshness_score": 0-100,
      "freshness_label": "fresh|acceptable|stale|indeterminate",
      "staleness_risk": "high|medium|low|none",
      "reasoning": "Brief explanation of the score, mentioning date comparison and volatility factors.",
      "should_use": true or false,
      "warning": "If should_use is false, provide a specific warning. Otherwise null."
    }
  ],
  "overall_assessment": "Brief summary of whether the quote set is temporally adequate for the query."
}

## CONSTRAINTS
- If a source_date is missing or unparseable, set freshness_score to 0, freshness_label to "indeterminate", and staleness_risk to "high".
- If the query has no time sensitivity (query_time_sensitivity is "none"), all quotes with valid dates should score at least 70 unless they contain explicit outdated language.
- Do not penalize a quote for staleness if the query is explicitly about a historical period that matches the source_date.
- For financial, medical, legal, or regulatory queries, apply stricter freshness thresholds: quotes older than 90 days relative to the reference date should be flagged as "stale" unless the query is explicitly historical.
- If staleness_risk is "high," should_use must be false.

## EXAMPLES
Query: "What is the current price of Apple stock?"
Reference Date: 2025-01-15
Quote source_date: 2024-06-01
Assessment: freshness_score=10, freshness_label="stale", staleness_risk="high", should_use=false, warning="Stock prices are highly volatile. This quote is over 7 months old and cannot represent the current price."

Query: "What were the key findings of the 2019 IPCC report?"
Reference Date: 2025-01-15
Quote source_date: 2019-09-01
Assessment: freshness_score=95, freshness_label="fresh", staleness_risk="low", should_use=true, warning=null
Reasoning: The query is about a specific historical report. The quote date matches the report publication period exactly.

Adapt this template by adjusting the [CONSTRAINTS] section to match your domain's freshness thresholds. For news applications, you might reduce the acceptable window to 24-48 hours. For legal research, you might extend it but add jurisdiction-specific staleness rules. The output schema is designed to be machine-readable for downstream filtering: use the should_use boolean to automatically exclude stale quotes from your evidence set before answer generation. Always test with known stale-evidence scenarios from your domain to calibrate the scoring thresholds.

The Quote Freshness and Temporal Relevance Prompt is designed to sit inside a retrieval-augmented generation (RAG) pipeline or evidence-ranking service, not as a standalone chatbot. Its job is to receive a set of extracted quotes—each with source metadata including a publication or revision date—and a time-sensitive query, then return a structured freshness assessment. The output includes a freshness score, a staleness warning flag, and a temporal relevance justification. This harness must enforce that every quote carries a date before the prompt is invoked; missing dates should cause the record to be rejected or routed to a separate date-extraction step before freshness evaluation.

Wire the prompt into your application as a post-retrieval, pre-synthesis filter. After your retrieval system returns candidate quotes and before those quotes are passed to an answer-generation or citation-rendering step, insert this freshness check. The application should call the prompt with a batch of quotes (each containing quote_text, source_date, and source_id) and the user's original query. Parse the structured output—a JSON array of freshness assessments—and apply a configurable threshold: quotes with a freshness_score below your cutoff or with staleness_warning: true should be dropped or demoted before downstream use. For high-stakes domains such as financial intelligence or regulatory monitoring, log every freshness decision with the quote ID, score, and justification for auditability. If the model returns malformed JSON, implement a retry with the error message injected into the prompt context, and escalate to a human reviewer after two failed attempts.

Model choice matters here. Use a model with strong date-reasoning capabilities and reliable JSON output—GPT-4o, Claude 3.5 Sonnet, or equivalent. Avoid smaller models that may mishandle relative date expressions like 'last quarter' or 'recently.' For latency-sensitive pipelines, consider caching freshness scores for quotes whose source dates haven't changed, and only re-evaluate when new quotes enter the system. Do not use this prompt to make automated trading, medical, or safety-critical decisions without human review. The freshness score is a signal, not a verdict. Pair it with source authority checks and content-faithfulness verification before any quote reaches a user-facing answer.