How internal search actually works, why wikis rot, how to structure a runbook for semantic retrieval, and why source citation matters more than answer fluency. Written by the team building Knowharbor.
Keyword search returns pages that contain your words. Embedding-based semantic search returns the answer to your question — even when the document uses completely different vocabulary.
New hires spend their first weeks asking questions that have written answers somewhere in Confluence or Google Drive. The problem isn't missing documentation — it's that the retrieval layer doesn't exist.
When the same two engineers handle every Tier 2 escalation, the instinct is to blame process or culture. But the root cause is structural — there's no knowledge retrieval layer between the ticket and the expert.
Most runbooks are written for the person who already knows the procedure. For semantic search to find the right runbook, it needs to be structured for the person asking the question for the first time.
A RAG tool that answers without showing the source document is a liability in any internal knowledge context. Source citation is the mechanism that keeps answers verifiable — and keeps your team from acting on hallucinated procedures.
A better search bar doesn't fix stale content. But stale content doesn't mean the wiki is useless — it means the retrieval layer needs to surface the best available answer and flag what needs updating.