Knowledge base

An assistant answering from your documents is exactly as good as your documents, and this is where most disappointments actually originate. The common experience is not that the model is stupid; it is that three versions of the refund policy exist in three places, the assistant found all of them, and answered from whichever happened to rank highest. That is your filing being described accurately for the first time, and it is fixable in a way a model limitation is not.

The defining act is deciding what is in it. Which documents, which versions, whose word counts as final, and what happens to the things nobody wants to delete. None of that is a technical question, and skipping it is why so many of these projects stall at the point where somebody has to choose between two policies with the same title.

Contradiction is the biggest practical problem and the least discussed. Where two documents disagree, retrieval surfaces whichever sits closest to the question, and the answer is written confidently with no indication that a conflicting statement existed. You will not be told there was a disagreement. You will simply get one side of it.

Currency has to be managed deliberately, because nothing expires by itself. A superseded document that was never deleted goes on answering questions indefinitely, and the older it is the more likely it is to be phrased with the confidence of something that was once true. An owner and a review date do more for answer quality here than any change of tooling.

Structure helps considerably more than people expect. Clear headings, a date, sections short enough to stand alone, and an explicit line saying what a document replaces all make the difference between a passage that retrieves usefully and one that arrives without the context that made it meaningful. Writing for retrieval is a small discipline with a large effect.

Scope and permission decide who may ask what. One base serving everybody is simple and means every answer is drawn from material everybody may see, which is usually too restrictive. Separating by department is generally easier to run than per-document rules, and either way the checks belong at the point material is retrieved rather than in the wording of the request.

• Load a deliberately chosen handful of documents into a notebook tool and notice how differently it behaves from a general assistant on the same questions.

  NotebookLM

• Ask a question of a notes workspace and get an answer from pages your own team wrote, which is a knowledge base you have been building for years without calling it one.

  Notion AI

• Search across a whole company's systems and see what happens when the same rule is stated in four places by four teams.

  Glean

What should go in it?

The material somebody would reach for to answer the questions you actually get, in its current version only. Start narrow, with one subject and its genuinely authoritative documents, and expand once that works. Everything you add is another candidate answer, which is a reason to be reluctant rather than generous.

Our documents contradict each other. What happens?

You get one of them, chosen by whichever sat closest to the question, presented without any hint that the other exists. This is a frequent cause of a confidently wrong answer from an otherwise well-built system, and the fix is editorial: decide which is current and remove or clearly mark the other.

Who should own it?

The team whose work the documents describe, not whoever built the assistant. Ownership means deciding what is current, reviewing on a schedule and arbitrating when two versions disagree. An unowned base degrades quietly, and the first sign is usually a customer being told something that stopped being true a year ago.

How do we keep it current?

Give every document an owner and a review date, and make deleting the superseded version part of publishing the new one. That is unglamorous and it is the whole answer. No amount of retrieval sophistication compensates for a shelf where the old edition is still sitting next to the new one.

• It is the defined set of material an assistant may answer from, and defining it is the actual work.
• Contradictory documents produce a confident answer from one side with no sign the other existed.
• Nothing expires on its own, so superseded material keeps answering until somebody removes it.
• Writing for retrieval, with dates, headings and self-contained sections, has a large effect.
• This is an editorial project with a technical component, and staffing it the other way round goes badly.

• NotebookLM

  A small, deliberately chosen shelf, which is the idea at a size you can hold.

• Notion AI

  A knowledge base most teams have been building without naming it one.

• Glean

  Company-wide, where the same rule stated four times becomes visible.