Freshness is one of the most visible trust signals in RAG. Users notice stale answers quickly, especially for policies, pricing, and operational procedures. Freshness evaluation is how you turn "it feels outdated" into measurable signals and release gates.
Define what freshness means per domain
Freshness is not one number. Define freshness targets by content domain (see freshness architecture):
- Policy and compliance content: strict freshness requirements.
- Product documentation: moderate.
- Evergreen knowledge: lower urgency.
Measure source age in answers
A practical metric is "answer age": the age of the newest (or primary) cited source. If answers cite sources older than your freshness target, you have a trust problem even if the text is "correct".
To measure this, your citations need metadata: effective date, last updated, and ingestion time (see metadata strategy and structured citations).
Use golden queries for freshness-sensitive workflows
Build a set of golden queries that should surface the latest policy or procedure and run them continuously. Alert when expected sources are missing or when older sources dominate (see synthetic monitoring).
Detect stale-source dominance
Beyond single answers, watch for patterns:
- High retrieval volume from sources older than the freshness target.
- Domains with ingestion lag exceeding the SLA.
- Frequent conflicts between old and new sources (see canonical sources).
Connect evaluation to operational levers
When freshness degrades, teams need fast levers:
- Re-ingest critical domains and process tombstones (see deletion workflows).
- Temporarily filter low-trust or outdated sources (see source quality scoring).
- Pause risky releases and stabilise ingestion (see change freeze).
Make freshness part of user trust
Where appropriate, surface freshness in the UI: show effective dates on citations and warn users when sources are old. This turns freshness into an explicit trust signal rather than a surprise (see user transparency).
Freshness evaluation is not just for data teams. It is a reliability control that protects user trust and reduces governance risk.