Knowledge is where you give the assistant your own material. It is the difference between an AI that knows about business generally and one that knows about yours.
What it does
Material you add is broken into chunks and indexed, so that when somebody asks a question the relevant pieces are found and used in the answer.
That mechanism matters for one practical reason: the assistant retrieves the parts of a document, not the document. A well-structured page with clear headings gets found precisely. A forty-page PDF with no structure gets found vaguely, and the answer is vague to match.
Scoped by HQ
Knowledge is filed against the area it belongs to, across every HQ in the platform.
A question asked in Finance searches Finance material first. The same question in Sales searches Sales.
Scope it properly when you add it. Material filed everywhere is retrieved for questions it has nothing to do with, and the effect is not neutral: it crowds out the material that was actually relevant. A refund policy that surfaces during a sales question makes every sales answer slightly worse.
Some things genuinely are global. Company policies, the way you refer to your own products, how you handle escalation. Those belong everywhere. Most things do not.
What to add first
In order of how much difference it makes.
Your own vocabulary. What you call your products, your stages, your teams, your customers. This is the highest-value material by a distance, because without it the assistant uses generic industry language and every answer reads as though it came from somebody who has not worked here.
Policies and rules. Refunds, escalation, approval thresholds, notice periods. Things people ask about constantly and look up in a document nobody can find.
How you actually do things. The process as followed, not the process as written down three years ago.
Reference material. Specifications, pricing rules, terms.
What not to add
Anything sensitive that people should not see. The assistant respects permissions on your records, but material you upload here is knowledge for answering questions. If something should only be seen by three people, it does not belong in a general knowledge base.
Anything out of date. Wrong material is worse than missing material, because a missing answer prompts somebody to check and a wrong one does not.
Everything. The instinct to upload the whole shared drive produces an assistant that retrieves plausibly-related noise for every question. Curation is the work here, not volume.
Structure beats volume
The single most useful thing to understand about adding material.
Because retrieval works on pieces rather than whole documents, how a document is organised decides how well it is found.
Headings that name the question. A section headed "Refund window" is retrieved for a refund question. The same content under "Section 4.2" is not.
One subject per section. A section covering refunds, exchanges and warranty gets retrieved for all three and answers none of them cleanly.
Say the thing, do not allude to it. "Customers may return items within 30 days" is retrievable. "The standard window applies" is not, because the standard window is stated somewhere else and the two pieces will not necessarily arrive together.
Spell out what your organisation calls things, at least once, in full. Documents written for people who already know the vocabulary teach the assistant nothing.
A short, well-structured page beats a long, thorough one. That is the opposite of how most internal documentation is written, which is why most internal documentation makes a poor knowledge base without editing.
Checking it worked
The page shows how many records are indexed, how many are embedded, and how many are still waiting.
A gap between indexed and embedded means the work is not finished. Material that has not been embedded is not retrievable, so it exists and does nothing. If that number stays above zero, something is not completing.
Use the search box before assuming. It shows what would actually be retrieved for a given question. Type a question somebody really asks and look at what comes back.
That check takes ten seconds and answers the only question that matters: is the right material being found. Uploading a document and assuming is how people end up believing the AI is bad when it simply never saw the file.
When answers are wrong
Almost always a knowledge problem rather than a model problem.
Search for the question first. If nothing relevant comes back, the material is missing, wrongly scoped, or not embedded.
If the wrong thing comes back, you usually have two documents saying different things, and the older one is winning. Remove the old one rather than adding a third.
If the right thing comes back and the answer is still wrong, then look at the model. That is the rarer case, and checking it first wastes an afternoon.
Who can add material
Adding knowledge is an administrative act, and it should be.
Material added here is used to answer questions for everybody, so a document uploaded by one person shapes what the whole organisation is told. That is not a permission to hand out widely.
Have one owner per area. Somebody in Finance who is responsible for what the assistant says about Finance. Diffuse ownership produces contradictory material that nobody notices adding.
Keeping it current
Knowledge decays quietly, because nothing breaks when a policy changes and nobody updates the copy the AI is reading.
Re-check after any policy change. The change is not real until the assistant knows about it, and somebody will ask the assistant before they read the announcement.
Remove rather than supersede. Two versions of a policy in the index is a coin toss on every answer.
Review what is there once a quarter. Less often than that and you will find material three years stale being quoted confidently to a customer.
What it costs to keep
Indexing material consumes a little, and keeping it current consumes almost nothing.
It is not a reason to hold back. The volume of material a business has that is worth the assistant knowing is small compared with what it produces daily, and the cost of an assistant that does not know your refund policy is much larger.
Related
Persona and guardrails controls how it sounds; this controls what it knows. They are separate problems, and confusing them is why some organisations keep adjusting tone to fix accuracy.
Watching AI usage shows what is being asked, which is the best available guide to what to add next.
Did this answer your question?
No, ask a person