All resources
Data & RAG23 March 2026·5 min read

Stale answers are worse than no answers: keeping a knowledge base honest over time

A RAG system launched on day one is accurate. The same system, untouched, six months later is quietly serving last year's policy with this year's confidence.

The uncomfortable truth about retrieval systems is that launch-day accuracy tells you nothing about month-six accuracy. Documents change, policies update, product specs shift — and unless someone designed for that, the index keeps serving the version it indexed at launch, with exactly the same confident tone it used on day one. A stale answer delivered with total confidence is worse than no answer at all, because the user has no signal to doubt it.

Where staleness creeps in

  • Source documents get updated in place but the index isn't resynced on any schedule, so the retrieval layer serves the pre-update version indefinitely.
  • Deprecated documents never get removed — old policy PDFs sit in the same folder as their replacements, and the system has no way to know which one is authoritative.
  • Nobody owns the sources — when no person is accountable for a given document category, updates happen inconsistently and the knowledge base drifts silently out of sync with reality.

What keeps a knowledge base honest

  • Scheduled resync — an automated pipeline that re-indexes on a cadence matched to how often the source actually changes, not a one-time launch job.
  • Visible provenance — every answer surfaces which document and version it came from, with a 'last verified' date, so a user can judge freshness themselves.
  • Owned sources — each document category has a named owner responsible for flagging updates and removals, not an assumption that content maintains itself.

Design the resync pipeline in the same sprint as the initial index, not as a follow-up ticket. 'We'll add freshness later' is how six-month-old policy documents end up answering compliance questions with complete confidence.

Why this matters more than it sounds

For anything touching policy, pricing, or compliance, a confidently wrong answer isn't a minor UX issue — it's the kind of thing that shows up in a customer complaint or an audit finding months later, traced back to a document nobody remembered to update.

Related capability

This is what our Custom LLM Integration practice is built around.

See how it's scoped