Source Currency Audit Prompt for Knowledge Bases

This prompt is designed for knowledge base maintenance teams and AI operations engineers who need a systematic, automated way to prevent stale information from degrading the quality of RAG applications, internal search, or customer-facing assistants. The core job-to-be-done is auditing a document corpus for currency, identifying entries that require updates based on age, domain-specific decay rates, and the availability of newer sources. The output is a prioritized audit report with refresh recommendations, enabling teams to move from reactive firefighting to scheduled, proactive content maintenance.

Use this prompt when you manage a knowledge base where information has a predictable shelf-life—such as product documentation, regulatory filings, financial reports, or technical runbooks—and you need to schedule regular audits. The prompt requires a structured input payload containing document metadata (publication date, last-updated date, document type, and domain), a set of domain-specific freshness thresholds, and an optional list of newer candidate sources. It is designed to be wired into a scheduled automation harness, such as a weekly CI/CD job or a cron-triggered cloud function, that feeds the prompt with a batch of documents and processes the structured audit report.

Do not use this prompt for real-time freshness checks on live data streams, for evaluating the factual accuracy of document content, or for making automated deletion or archival decisions without human review. The prompt assesses currency based on temporal signals and source availability, not on content correctness. For high-risk domains such as healthcare or legal compliance, always route the prioritized audit report to a human reviewer before taking action. If your knowledge base contains documents without reliable publication or last-updated metadata, pair this prompt with a Publication Date Extraction Prompt to normalize timestamps before running the audit.

This prompt is the core instruction set for a Source Currency Audit. It is designed to be dropped into an orchestration layer where variables like the knowledge base snapshot, domain-specific decay rules, and the output schema are populated at runtime. The template uses square-bracket placeholders exclusively to prevent conflicts with other templating engines. Before pasting, ensure you have a clear definition of your domain's decay rates and a list of any newer source collections you want the auditor to compare against.

textCopy
You are a Knowledge Base Currency Auditor. Your task is to analyze a snapshot of a knowledge base and identify entries that are outdated or at risk of becoming stale.

# INPUTS
- Knowledge Base Snapshot: [KNOWLEDGE_BASE_SNAPSHOT]
- Domain Decay Rules: [DOMAIN_DECAY_RULES]
- Newer Source Collections (Optional): [NEWER_SOURCE_COLLECTIONS]
- Current Date: [CURRENT_DATE]

# OUTPUT SCHEMA
You must produce a JSON object with the following structure:
{
  "audit_report": {
    "audit_timestamp": "string (ISO 8601)",
    "entries_analyzed": "integer",
    "stale_entries": [
      {
        "entry_id": "string",
        "current_content_summary": "string",
        "staleness_reason": "string (e.g., 'Exceeded domain decay window', 'Superseded by newer source')",
        "recommended_action": "string (e.g., 'REFRESH', 'ARCHIVE', 'REVIEW')",
        "urgency": "string (e.g., 'HIGH', 'MEDIUM', 'LOW')",
        "suggested_replacement_source": "string | null"
      }
    ],
    "healthy_entries_count": "integer"
  }
}

# CONSTRAINTS
- An entry is stale if its last-updated date is older than the domain-specific decay window defined in [DOMAIN_DECAY_RULES].
- If [NEWER_SOURCE_COLLECTIONS] is provided, also flag entries that are superseded by more recent, authoritative information from those collections.
- Base your analysis strictly on the provided inputs. Do not invent external facts.
- If an entry's last-updated date is missing, classify its urgency as 'HIGH' and its recommended action as 'REVIEW'.
- Prioritize entries that are both stale and frequently accessed or critical to core operations, if such metadata is available in the snapshot.

To adapt this template, replace the placeholders with your actual data. [KNOWLEDGE_BASE_SNAPSHOT] should be a structured list of entries with their IDs, content summaries, and last-updated timestamps. [DOMAIN_DECAY_RULES] is a critical input; define it as a clear set of rules, such as 'Financial regulations: 90 days; Technical documentation: 365 days; News articles: 7 days.' The [NEWER_SOURCE_COLLECTIONS] placeholder is optional but powerful—use it to provide a set of documents from a recent crawl or a premium data feed that might supersede your existing knowledge base. The next step is to wire this prompt into a scheduled job that validates the JSON output and feeds it into your content refresh pipeline.

The Source Currency Audit Prompt is designed to run as a batch or scheduled job, not a real-time user-facing endpoint. The typical integration pattern is a cron-triggered script or a step in a data pipeline (Airflow, Prefect, Dagster) that iterates over knowledge base entries, sends each to the model with the audit prompt, collects structured audit reports, and writes results to a review queue or dashboard. Because the prompt produces a structured JSON report with prioritized refresh recommendations, the harness should validate the output schema before accepting any result. A malformed or hallucinated audit entry can cause downstream automation to skip a genuinely stale document or flag a current one for unnecessary review.

Start by defining the input batch. For a knowledge base of 10,000 articles, you might process 100 per run to stay within rate limits and cost targets. Each article needs its [DOCUMENT_METADATA] (title, URL, last_modified date, content type) and [DOCUMENT_CONTENT] (full text or a representative chunk). The prompt template expects these as structured inputs. On the model side, prefer a model with strong structured output support (GPT-4o, Claude 3.5 Sonnet, or Gemini 1.5 Pro) and enable JSON mode or structured outputs with the schema defined in the prompt's [OUTPUT_SCHEMA] block. Set temperature=0 to minimize variance across audit runs. Implement a retry wrapper: if the model returns invalid JSON, a missing required field, or a confidence score below your threshold (e.g., <0.7), retry once with the same input. If the second attempt also fails, log the failure and flag the entry for manual review rather than silently accepting a bad audit result.

The audit report output should be written to a persistent store (database table, blob storage, or a review tool like Retool or a custom dashboard). Each record must include the original document identifier, the audit timestamp, the model's freshness score, the recommended action (keep, review, replace, archive), the decay rate justification, and any suggested replacement sources. Build a validation layer that checks: (1) the recommended_action is one of the allowed enum values, (2) the freshness_score is a float between 0 and 1, (3) suggested_replacements is an array even if empty, and (4) the audit_timestamp is within the current run window. Reject and retry any output that fails these checks. For high-stakes knowledge bases (compliance, legal, financial), route all replace and archive recommendations to a human review queue before any automated action is taken. Never auto-delete or auto-archive content based solely on model output.

For observability, log every audit run with: document ID, model used, prompt version, latency, token count, validation pass/fail, retry count, and the final action taken. This creates an audit trail for the audit process itself. If you run this weekly, compare week-over-week results to detect drift—if a document flips from 'current' to 'stale' and back without content changes, your prompt or model behavior may need investigation. Set up alerts for sudden spikes in 'stale' classifications that could indicate a model regression or a real content decay event. Finally, version your prompt template in a prompt registry or version-controlled config file. When you update the decay-rate thresholds or the output schema, run the new version against a golden set of 50-100 documents with known freshness states to verify the audit behavior hasn't degraded before promoting to production.