RAG Knowledge Half-Life Calculator

Translate freshness audits into an exponential decay model so you can quantify how quickly retrieval-augmented generation answers go stale and schedule refresh work before coverage slips.

Number of days between the baseline snapshot and the evaluation you are using to measure decay.
Total retrieval-augmented responses reviewed for freshness during the observation window.
Number of reviewed answers flagged as outdated, inaccurate, or missing critical context.
Defaults to 80%. The share of answers you want to remain current before the next refresh.
Defaults to 100%. Adjusts the stale rate upward when less than full coverage is achieved.

Validate decay assumptions with periodic back-testing and incorporate human review for high-risk knowledge domains before rolling changes to production assistants.

Examples

  • 30-day window, 120 answers reviewed, 18 stale, 80% target freshness, 60% coverage ⇒ Knowledge half-life: 54.83 days (decay rate 0.0126 per day). Plan refreshes every 25.37 days to keep at least 80.00% of answers current.
  • 14-day window, 90 answers reviewed, 6 stale, optional fields blank ⇒ Knowledge half-life: 162.03 days (decay rate 0.0043 per day). Plan refreshes every 72.27 days to keep at least 80.00% of answers current.

FAQ

Why is exponential decay appropriate for knowledge freshness?

Most retrieval corpora accrue staleness gradually as policies, product specs, and source documents change. Assuming a constant proportional decay rate approximates this behaviour and lets you back-solve half-life from any observation window.

How should I choose the target freshness threshold?

Set the threshold to the minimum share of answers that must remain current for your service-level objective. Regulated workflows may need 90%+; internal assistants can often tolerate 70–80% before triggering bulk updates.

What if coverage is below 100%?

Enter the approximate share of traffic touched by evaluations. The calculator inflates the observed stale rate so you do not underestimate decay when sampling is sparse.

Can I convert the half-life into weekly decay?

Yes. Multiply the daily decay rate shown in the output by seven to obtain an approximate weekly rate, or re-run the calculator with observation days expressed in weeks if that matches your reporting cadence.

Additional Information

  • Stale fraction derives from the share of evaluated answers flagged during the observation window, scaled by evaluation coverage to avoid underestimating decay.
  • The calculator assumes exponential decay, translating observed staleness into a daily decay constant and half-life.
  • Refresh cadence indicates when to trigger knowledge base updates before freshness falls below the target threshold.

Related calculators