RAG that stays fresh: freshness SLAs beat bigger indexes

A large index of stale documents is worse than a small index of live ones. Freshness is a product requirement, not a backend chore.

Name the sources and the clock

Every corpus should declare where it comes from and how old it is allowed to be. Policy docs may tolerate a day. Inventory and pricing may not tolerate an hour. Put that SLA on the agent card, next to latency.

• Connector health is part of agent health
• Citations must include retrieved-at timestamps
• If retrieval confidence is low, say so and route to a human
• Re-embed on change events, not on a hopeful nightly job

An agent that cannot tell you when it last read the source is not knowledgeable. It is confident.

Measure retrieval, not only answers

Track hit rate, citation click-through, and “correct source, wrong span.” Those metrics move enablement forward. Golden-answer evals alone hide rotting knowledge.