Picture an organisation that launched a copilot whose answers are sometimes right and sometimes stale. The first instinct is to swap the model — but the root cause sits elsewhere: the policy is in one folder, its new version in an email, and the exceptions in one expert’s head.
What a knowledge spine is
A knowledge spine is a layer that indexes scattered sources in one place — each with an owner, a validity date and an access level. Instead of searching everything, the copilot reads from this layer and returns every answer with its source. Technically it’s RAG; organisationally it’s the discipline behind it.
Where to start
Pick one small, high-volume area — HR questions or support procedures. List its sources, retire old versions and give every document an owner. Then chunk and index the documents, and keep a few dozen real questions as a test set.
The common mistake
The common mistake is feeding every document to the copilot at once. The result is answers quoting obsolete files. A spine should grow, not be filled overnight — and whenever a policy changes, the index must change too, which a simple automated workflow can handle.
In our enterprise LLM projects we follow exactly this path: one area, one metric on real questions, and expansion once the results are in.