Knowledge base engineering

A vector index is not a knowledge base. It is one representation of one moment of a corpus, with no owner, no lifecycle, no provenance, and no idea what supersedes what. Most "knowledge bases" built in the last few years are exactly this, which is why they work in the demo and decay in production: content changes, nobody is accountable, a second team builds a second index over the same documents, and within a year the organization has several inconsistent memories of itself.

My position is that a knowledge base is an engineered asset: a governed set of representations of source content, each with provenance and context, owned by the people who understand it, operated on a shared platform with explicit contracts, and managed through a lifecycle from ingestion to retirement. Knowledge base engineering is the discipline of building and running that asset. It is a distinct practice — not data engineering, not machine learning, not content management — and most organizations do not yet have anyone doing it.

This piece is about what the discipline consists of.