RAG (retrieval-augmented generation) makes your company's documents answerable: employees ask questions, the system retrieves the relevant passages and drafts a cited answer. But the technology is the easy half, knowledge ops, deciding what's in the index, who owns it and when it expires, is what separates a trusted answer machine from a confident liar.
Knowledge ops: the part that isn't RAG
- Curate in, don't dump in: start with the 200 documents people actually need, policies, runbooks, product docs, pricing, not the whole drive.
- Ownership: every indexed source has a named owner responsible for its accuracy.
- Freshness rules: docs expire; expired docs leave the index or get flagged in answers.
- Conflict resolution: when two docs disagree, that's a knowledge bug, route it to the owners instead of letting the model pick.
- Access control: the index must respect permissions, RAG that leaks the salary file is a career-ending deployment.
Measuring whether it works
- Collect 50 real questions from support, onboarding and internal chat.
- Grade answers weekly: correct, cited, current. Track the score like uptime.
- Log 'no good answer' cases, they're your content roadmap.
- Watch deflection honestly: did people stop asking colleagues, or stop trusting the tool?
A rollout sequence that builds trust
- 1One high-pain domain first (support macros, HR policy, or engineering runbooks).
- 2Pilot with a friendly team; grade every answer for two weeks.
- 3Fix the corpus, not the prompt, most wrong answers trace to wrong or missing docs.
- 4Expand domain by domain, each with an owner, never 'index everything' in one step.
