← Back to Blog

Conversational AI Chatbot Knowledge Management: A Guide

Conversational AI chatbot knowledge management is the work of deciding what your chatbot knows, keeping that knowledge current as your product changes, and closing the gaps real conversations expose. It runs as three loops: sources (what goes into the knowledge base), gaps (questions it couldn’t answer well) and review (a person approving what goes live). Total Chat grounds its answers in a searchable knowledge base. You seed it with your own articles, the URL crawler and the codebase scanner, and on Starter and up the assistant drafts new articles from conversations that exposed a gap. Those drafts don’t publish until you approve them.

Last updated: October 2026

The part that changed is who writes the first draft

For years, chatbot knowledge management meant someone on the team reading transcripts on a Friday afternoon, spotting the questions the bot fumbled, and writing help articles to cover them. Most weeks that person had better things to do. No wonder the knowledge base froze at launch while the product kept shipping. If you’ve suspected your chatbot is still describing last quarter’s onboarding, it probably is.

The shift isn’t that the chatbot got smarter on its own. It didn’t, and you shouldn’t want it to. What changed is that the conversation exposing the gap can now produce the draft that fills it. Your job moves from writing to reviewing, and reviewing is the job that actually gets done.

If you want the background on how these assistants work in general, start with What Is a Conversational AI Chatbot? This post assumes you have one and want its answers to stay right.

Loop 1: sources, or what the chatbot is allowed to know

A chatbot is only as current as the material it searches. In Total Chat, that material comes from three places, and they do different jobs.

Articles you write. You create articles through the knowledge base API, and the dashboard is where you review, edit, approve or reject drafts. An article can carry a page route, keywords and a list of the common questions it answers, and search ranks articles tied to the user’s current page higher. Hand-written articles are where your precise steps, troubleshooting and plan details belong.

The URL crawler. Point it at a live URL and it follows internal links on the same domain, reads each page’s title, headings and visible text, and generates 3 to 8 draft articles. It’s on every plan, including Free. It’s the fastest way to get from an empty knowledge base to a usable one, but those drafts are a starting point. Edit them to add accurate step-by-step instructions, common troubleshooting scenarios and pricing information before you approve them.

The codebase scanner. It builds a feature map of a Next.js app’s routes, features and UI selectors (run with npx @totalchat/cli scan --upload), which is what lets the logged-in assistant know which page, tab and modal a user is on and what they can do there. Each upload also writes a short article per feature category plus a navigation guide. Those publish straight away while you have room under your plan’s article cap, and each re-scan refreshes them. They list what exists, not how to use it, so they don’t replace your own articles. Only Next.js apps are supported right now.

One decision people skip: which articles pre-login visitors should see at all. Each article has a public visibility flag. Turn it off and the article only surfaces in logged-in chat, and the server enforces that, so a visitor can’t get around it. Use it for anything about cancellation, billing disputes, internal processes or admin workflows that you wouldn’t show an anonymous visitor on your marketing site.

Loop 2: gaps, or what it couldn’t answer

You find gaps in two ways: the assistant tells you, or you go looking.

On Starter and up, when a conversation reveals something the knowledge base doesn’t cover, the assistant drafts a knowledge base article from it. The draft lands in the dashboard’s pending drafts, and you get an email about it. You approve it, edit it or reject it. It does not publish itself.

The dashboard’s conversation metrics and top questions are the other route. I’d read them weekly even with drafting turned on, because the most useful gaps aren’t always single unanswered questions. Sometimes it’s a handful of slightly different questions that all mean “your onboarding is confusing,” and no single article fixes that.

Not every gap is a knowledge problem, either. When a user says something is broken, the assistant asks a short follow-up and files a structured bug report with the page address, browser and recent JavaScript errors. That’s on every plan. On Pro and up, those reports can also be created in Zendesk, Freshdesk, Jira Cloud, Linear, Help Scout or a signed webhook. A bug filed as a bug is better than a help article explaining around it.

Loop 3: review, or what goes live

Here’s my opinion, plainly: a chatbot that publishes its own articles is the wrong default. A draft that sounds right and is wrong does more damage than a gap, because the assistant will repeat it with total confidence to every user who asks.

So treat the knowledge base like code. Changes are proposed, a person reviews them, and only then do they ship. In Total Chat, crawler drafts and conversation drafts both wait for approval. The scanner’s feature lists are the one exception, because they come straight from your code rather than from anything the assistant wrote. Published articles are what count toward your plan’s cap. Drafts awaiting approval don’t.

The caps by plan are 20 published articles on Free, 100 on Starter, 500 on Pro and unlimited on Business. The full table is under plans and limits. If you’re close to a cap, that’s a useful forcing function. Merge overlapping articles before you buy room for more.

Keeping it current when the product ships

Knowledge goes stale on deploy day, not on a schedule. Two habits help.

First, re-run the scanner when features change. It can run as a GitHub Actions step on every push to main, and a freshness check reports whether the uploaded map is fresh, getting old or stale. You can also notify Total Chat from your CI/CD pipeline on each deploy, including the commit SHA, so the freshness check can tell when your deployed code no longer matches the scanned map.

Second, re-crawl or edit articles when your marketing pages or docs change. The crawler only sees what’s published, so a pricing change on your site isn’t in the chatbot’s answers until an article says so.

You don’t need a dedicated knowledge manager for this. A short weekly review covers most teams: clear the draft queue, skim the top questions, retire one stale article. It’s boring work, and boring is fine.

Where this approach has limits

Gap drafting isn’t on the Free plan. The scanner reads Next.js apps only. And the pre-login widget doesn’t use the visitor’s current page in its answers, so on your marketing site the knowledge base and the visibility flag do all the work. That’s also where most first conversations happen. Lead capture runs on every plan there, and Conversational AI Chatbot for Lead Generation covers how that side works. Setup details for everything above are in the docs.

Frequently Asked Questions

What is conversational AI chatbot knowledge management?

It’s the practice of controlling what a chatbot knows and keeping it accurate over time: choosing sources, finding the questions it can’t answer, and reviewing changes before they go live. In Total Chat, answers are grounded in a knowledge base you seed with your own articles, the URL crawler and the codebase scanner, and drafts written from conversations or crawls wait for your approval.

Does Total Chat update its knowledge base on its own?

Not with anything the assistant writes. On Starter and up, when a conversation exposes a gap, it drafts a new article, and you approve, edit or reject it in the dashboard first. URL crawler drafts wait for review too. The exception is the codebase scanner: each upload publishes a short feature list per category, built from your code, while you have room under your plan’s article cap.

How do I keep chatbot answers current when my product changes?

Re-run the codebase scanner after feature changes. It can run as a GitHub Actions step on every push to main for Next.js apps, it refreshes the scanner’s feature articles, and a freshness check shows whether the map is fresh or stale. For content changes, edit the affected articles or re-run the URL crawler against your updated pages, then approve the new drafts.

How many knowledge base articles can I publish?

It depends on the plan: 20 published articles on Free, 100 on Starter, 500 on Pro, and unlimited on Business and Enterprise. Only published articles count toward the cap, so drafts waiting for approval don’t use it up. The plans and limits table in the Total Chat docs lists the current caps alongside conversation and message limits.

Can I keep some articles out of the pre-login chat?

Yes. Each article has a public visibility setting. Turn it off and the article only surfaces in logged-in chat, never to anonymous visitors on your marketing site. The filter is enforced on the server. Use it for anything about cancellation, billing disputes, internal processes or admin workflows.

---