Train your AI agent on your knowledge base

An AI support agent has no opinions of its own. Everything it tells a customer comes from the knowledge base you give it, so the base is the real product you are building — not a setting you configure once, but an artifact you maintain.

Get the base right and the agent quotes your content back accurately. Get it wrong and the same agent states your mistakes with total confidence. What follows is how to build a base that answers well: where the content comes from, what makes a document answerable, how to keep it current, and how to let real questions tell you what to write next.

Where the knowledge comes from

There are four ways content gets into the base, and they are meant to be combined:

• Crawl a site. Point the crawler at your domain and it pulls in your public pages — help center, product pages, policies. Crawl the whole site, not just the homepage, so the agent sees the pages customers land on.
• Upload documents. Add PDFs and DOCX files directly: manuals, spec sheets, onboarding packets, the internal policy doc that never made it to the website.
• Paste raw text. For anything that lives in someone's head or a spreadsheet, paste it in as plain text. This is the fastest way to capture a question you answer every day but have never written down.
• Add FAQ pairs. Write the exact question and the exact answer. FAQ pairs are the most direct control you have over what the agent says, and they carry extra weight when the agent searches for a match.

The full set of sources lives on the knowledge base feature page. Once a source is added, you can inspect exactly what the agent learned from it, which is worth doing the first few times so you know what went in.

What makes a document answerable

Not every document helps. The agent retrieves passages and builds a reply from them, so a source is only as useful as it is easy to pull one clear answer out of. A few things separate a document that works from one that quietly hurts you.

• One topic per page. A page that covers returns, shipping, and warranty in three loose sections is hard to retrieve cleanly. Split it so each answer has a home.
• Real headings. Headings tell the retriever what a section is about. A wall of text with no structure forces it to guess where the relevant part starts and stops.
• Plain, current facts. Stale prices, retired plans, and last year's return window are worse than no answer, because the agent will state them plainly. If a number can change, keep it in one place you update.
• No contradictions. If two documents disagree about the refund window, the agent can surface either one. Reconcile duplicates before you add them, not after a customer quotes the wrong one back to you.

The goal is a grounded answer: a reply the agent builds from your content and can trace back to the source, rather than one it invents. Clean, well-structured sources are what make that grounding possible in the first place.

Keep it current with scheduled re-crawl

Content drifts. You change a policy, retire a plan, update a phone number. The base still holds the old version until you refresh it. A scheduled re-crawl re-reads your site on a set cadence, so edits you make on your website flow into the agent's knowledge without anyone re-uploading anything.

For documents and pasted text the rule is simpler: when the source of truth changes, replace the source. Either way, the point is the same. A base you set up once and never revisit slowly becomes a base you can't trust, and a support agent you can't trust is one you end up checking behind, which defeats the purpose.

Let the gaps tell you what to write next

You will never predict every question a customer asks. The useful signal is the set of questions the agent couldn't answer: the ones where it correctly said "I don't know" because nothing in the base covered them.

Knowledge-gap detection captures those misses and groups similar ones into themes, so instead of a raw log you get a short list of what is missing. A cluster of unanswered questions about international shipping is a page you haven't written yet. Turn one of those gaps into a new FAQ and it closes; the next person who asks the same thing gets a real answer instead of a shrug.

This is the loop that keeps the base alive: customers ask, gaps surface, you write, the base gets deeper. The questions your customers raise become your content roadmap, which means you spend your writing time on the things people genuinely need rather than the things you assumed they would.

A first-week sequence

You don't need everything in the base before you start. Work in this order:

1. 1Crawl your site. Get your public content in first. It is the widest coverage for the least effort.
2. 2Upload your canonical documents. Add the policies, manuals, and pricing that customers ask about but that the site covers poorly or not at all.
3. 3Add FAQs for your daily questions. Write the ten questions you answer most, with the answers you would give. These carry weight in retrieval and set the tone.
4. 4Test with real questions. Ask the agent the things customers ask, including the awkward edge cases, and read the answers critically rather than skimming them.
5. 5Turn on scheduled re-crawl. Set a cadence so your site content stays fresh on its own.
6. 6Review the gaps. At the end of the week, look at what the agent missed and write the two or three sources that would have answered it.

After that first week the work settles into a rhythm: a short weekly pass over the gaps, and an update whenever a real fact changes. Neither takes long once the base exists.

How Evoriqa fits in

Evoriqa handles the mechanics of all of this: crawling and re-crawling, document uploads, pasted text, FAQ pairs, and the hybrid retrieval that finds the best-matching passage and grounds each answer in it. When nothing in the base fits, the agent says "I don't know" and hands off rather than guessing, and that miss becomes a gap you can act on. The knowledge base is yours to shape; the platform's job is to keep it searchable, current, and honest about what it doesn't know.

To see it work against your own content, create a free workspace and point it at your site. No credit card, and the first crawl takes minutes.