How to Add Your Business Knowledge

Three ways in

The Grounded AI Agent & Knowledge Base answers from content you supply, so the first real task is supplying it. There are three routes, and each suits a different kind of material.

Paste text. The fastest option and the most underused. If the answer lives in an email you have sent forty times, paste it. No formatting to clean up, no file to find.

Upload files. PDF, DOCX, TXT, Markdown or CSV. Right for material that already exists as a document: service specifications, terms, price lists, product data exported as CSV.

Import a page by URL. Pulls the readable text of one public web page. Good for content that lives on your own site already, like an FAQ page or a service description.

What to add first

Start with what your team repeats. Open your sent folder, or think about the last twenty enquiries, and write down the questions that came up more than twice. That list is your first knowledge base, and it is almost always more useful than uploading your entire document library.

A practical starting set for most businesses:

• Opening hours, including anything seasonal
• Service or product descriptions, in customer language
• Coverage area or delivery terms
• Pricing structure, or how a quote is arrived at if you do not publish prices
• Lead times and what affects them
• Warranty, guarantee or returns terms
• The three things people always misunderstand about what you do

That last one is worth an evening on its own. Whatever your team spends most time correcting is exactly what the AI needs written down.

Processing, status and reindexing

After you add a source, it processes. The list shows status, so you can see when a source is ready rather than guessing why an answer has not changed yet.

You can edit source text directly in the workspace, which is the right move for a small correction, a changed price, a retired service. After an edit, reindex so retrieval picks up the new text. Sources can also be deleted when they stop being true, and deleting outdated material is as valuable as adding new material. Stale content that still retrieves well is worse than no content at all.

Write for retrieval, not for a brochure

Two habits make a noticeable difference to answer quality.

Use the customer’s words. If people say “callout fee” and your document says “attendance charge”, retrieval has a harder job. Include both.

Keep each source about one thing. A single document covering pricing, hours, coverage and terms retrieves worse than four short ones. Retrieval works on passages, so clean topical boundaries help it find the right passage.

Short question-and-answer formatting works particularly well, because it mirrors how the question will arrive.

The limits, stated plainly

Three boundaries to plan around rather than discover:

One page per URL import. There is no full-site crawl and no continuous website sync. If your FAQ page changes, reimport it or edit the source. Nothing updates itself.

Scanned PDFs need OCR first. Built-in OCR is not part of the product, so a scanned document with no text layer will not produce usable knowledge.

Quotas count sources and calculated pages. A knowledge page is calculated from 3,000 characters of extracted text, which will not always match one original PDF page. Internal search chunks are not what gets counted.

Test what you added

The fastest way to check your knowledge base is to interrogate it. Ask the chat the ten questions from your list. Anything it answers well is done. Anything it refuses is a missing source, and you now know exactly which one. Anything it answers wrongly is usually two sources disagreeing with each other, which is worth finding early.

Run that test again after any significant edit. It takes ten minutes and catches most of what would otherwise reach a customer.

Organising more than a handful of sources

Once you pass a dozen sources, structure starts to matter. A knowledge base is not a filing cabinet, nothing needs to be tidy for its own sake, but a few decisions make retrieval noticeably better.

Split by topic rather than by document. If your terms and conditions cover returns, delivery and warranty, three focused sources will retrieve better than one long one, because each passage sits closer to the question it answers. Conversely, do not shred a coherent explanation into fragments; a paragraph that only makes sense with its neighbours should stay with them.

Name sources so a colleague can audit them later. “Delivery and lead times, updated Sept” is a name that tells you when to check it. “Doc 4” is not.

Keep one source per fact. The most common cause of a confidently wrong answer is two sources that disagree: an old price list and a new one, both indexed and both retrievable. Whichever wins is essentially arbitrary. When you update something, update or delete the old version rather than adding alongside it.

A maintenance rhythm that survives a busy month

Knowledge bases decay quietly. Nothing breaks; the answers just drift further from what you would say today.

A workable rhythm is quarterly, plus an event trigger. Quarterly, reread the sources with a date, a price or a lead time in them. On an event such as a price change, a retired service or a revised policy, update the source the same week, while someone still remembers the detail.

The other half of maintenance is your transcripts. Refusals tell you what is missing; wrong answers usually tell you what is contradictory. Ten minutes reading conversations once a month will surface more genuine problems than any amount of speculative tidying.

One habit worth keeping

Whenever you answer a question you have answered before, ask whether it is in the knowledge base. If not, paste it in while the wording is fresh. Two minutes at the moment you notice beats an hour of retrospective auditing later, and it is how a knowledge base stays current without ever becoming a project.

Read next: Strict, Balanced and Open grounding modes compared, or how to reduce AI chatbot hallucinations.