Building a knowledge base your AI can read

A knowledge base for AI is not just another wiki

Many teams have already tried an internal wiki. People fill it enthusiastically for a month, then nobody opens it any more, and it quietly goes out of date.

A base read by AI assistants escapes that fate for a simple reason: it is used every day, without anyone having to open it. Every time someone asks their AI a question, the base is consulted. A useful document has an immediate effect, and so does a wrong one. The team therefore has a concrete reason to keep it up to date.

But the reader has changed, and that alters the way you write. A colleague reading a page guesses what is missing, knows who to call, and recognises an out-of-date version. An AI assistant reads what is written, and only that.

How an assistant reads your documents

To write usefully, you need to know what happens when your AI consults the base.

1. It searches. Starting from your question, it runs a search with a few words. The result depends on the words present in your documents and in their titles.
2. It receives excerpts. A few short passages, with the place they come from. On that basis it decides what to read.
3. It reads a part. It opens the most promising document and reads the relevant section, not necessarily the rest.
4. It answers, relying on what it has read.

Three writing rules follow directly.

• The words in the question must be in the document. If everyone says “quote” and your document talks about a “pricing proposal”, the search misses it.
• Each section must make sense on its own. The assistant may read section 4 without having read section 1. “As seen above” tells it nothing.
• A document must say whether it is valid. Between two versions that contradict each other, the AI cannot guess which one you apply.

What to write first

Don’t try to cover everything. Start with what you re-explain to your AI most often. In most teams, that comes to six documents.

You will find ready-to-adapt versions on the Templates page.

Writing a document an AI can make good use of

One subject per document

A document that covers invoicing, leave and onboarding new starters won’t be found properly for any of the three. Separate them. Three short documents are better than one long one.

A title that announces the content

“Procedure” says nothing. “Procedure: handling a customer complaint” can be found and chosen. Do the same for subheadings: “Response times” and not “Step 2”.

The summary at the top

Start with two or three sentences that say what the document is about, who it applies to and since when. This is often the passage that comes up in the search and persuades the assistant to read on.

Full sentences, with the reasons

Compare.

• “Discount: none in first year.”
• “We don’t grant a discount in the first year, because the price already includes support with getting started. From the second year, a discount of ten per cent is possible for a two-year commitment.”

The second version lets the AI answer a client, deal with a similar case, and explain the rule instead of reciting it.

Consistent terms

If you call the same thing “client”, “account” and “buyer” depending on the page, a human reader keeps up. An assistant may think these are three different things. Choose one word, and give the synonyms once in the “Who we are” document.

Examples

A rule about tone is easier to grasp with two examples, one good and one bad, than with three paragraphs of explanation. For a procedure, one real case handled from start to finish is worth a page of theory.

Special cases, written down

“Except for public contracts”, “unless the client is in a trial period”: exceptions are what people know without ever having noted them down. They are also where the AI goes wrong. Write them under the rule.

Organising: projects, links and status

Organise by project

Group together what is read together. One project for what is common to everyone, and one project per client, per assignment or per product. This division does three jobs: you can find your way around, the assistant searches within a relevant scope, and you grant access project by project.

Link documents to each other

A decision points to the meeting notes where it was made and to the study that justifies it. A procedure points to the template it uses. These links let the assistant, like a colleague, trace things back to the source when it needs to know more.

Say what is authoritative

This is the rule that pays off most and is applied least. Every document should carry a clear status.

• Authoritative: this is the reference, to be applied.
• Exploratory: a lead, thinking in progress, not a rule.
• Superseded: no longer applies, with a link to the document that replaces it.

Without this, the proposal abandoned in March has, for an AI, the same value as the decision made in June.

Importing what you have without importing the mess

You already have Word files, PDFs and spreadsheets. There is no need to rewrite everything, and no need to import everything.

1. Sort first. Set aside drafts, duplicates and anything that exists in more than two versions. Import only what is still true.
2. Import what is useful. Reference documents, templates, standard contracts, studies. The details of what is carried over for each format are in Importing your documents.
3. Mark the status on import. An old document that is useful as an archive must be flagged as such.
4. Rewrite only the essentials. The three or four most consulted documents deserve to be reworked according to the rules above. The rest can stay as it is.

Keeping the base up to date

A wrong base does more damage than an empty one, because the AI applies what it reads with confidence. Four habits are enough.

• Correct at the source, straight away. When a rule changes, you edit the document the same day, not “at the next update”.
• One owner per document. Not to write everything: to say yes or no when someone proposes a change.
• Let the AI propose. At the end of a meeting or a project, ask your assistant to record the decision or to propose an update to the procedure. You review it and approve it.
• Reread what is used most. Once a quarter, go back over the ten most consulted documents. That is where a mistake is costly.

Checking that it works

Write down ten questions that the team really does ask its AIs. Open a fresh conversation, without pasting anything, and ask them one by one.

• The answer is right and cites the right document: job done.
• The answer is vague: the document is missing, or it doesn’t use the words in the question.
• The answer is confidently wrong: an old document is lying around and isn’t marked as superseded.
• The assistant doesn’t consult the base: spell out in your instructions the cases in which it should search there.

Run the test again after each round of corrections. Ten questions are enough to spot most of the faults.

The seven mistakes to avoid

1. Importing the whole shared drive at once.
2. Letting contradictory documents sit side by side with no status.
3. Writing notes for yourself, in shorthand.
4. Putting ten subjects in a single document.
5. Giving the rule without the reason or the exceptions.
6. Forgetting to record decisions.
7. Writing once, and never correcting again.

Building this base with Linkevio

Linkevio is made for this. You write in an editor or import your Word, PDF, Excel and Markdown files. You organise by project, link documents to each other, and indicate for each one whether it is authoritative, exploratory or superseded.

You then connect the assistant you already use. It searches the projects you have access to, receives short excerpts with their location, reads the relevant part, and can write in the base if you ask it to. Every document keeps its versions, and every request from an assistant is recorded in a log.

To get started, follow the Getting started guide, pick one of the templates, and create your free workspace: it covers two people and one project, enough to write your first six documents and run the ten-question test.

Frequently asked questions

Do you have to write differently for an AI and for a human?

Hardly at all. A document that is clear to a new colleague is clear to an AI: one subject, a precise title, full sentences, reasons, examples. The only extra requirement is that each section makes sense on its own.

Which file format is read best?

Structured text, with titles and subheadings. A document written in an editor or in Markdown is ideal. Word files and text-based PDFs read well. A PDF made of scanned images, with no text, cannot be searched.

How many documents does it take to be useful?

Not many. Six well-written documents already change your assistant’s answers. Twenty up-to-date documents are better than two hundred of which half are out of date.

How long should a document be?

As long as it takes to cover one subject, rarely more than two or three pages. If a document goes beyond that, it probably contains two subjects.

Can the AI write the base for me?

It can help you put what you know into shape: you dictate, it structures. It cannot invent your methods or your decisions. Review by the person who knows remains essential.

What should we do with confidential documents?

Put them in a separate project, open only to the people concerned. An assistant only has access to its user’s projects. For more, see Sharing context across your team’s AI assistants.