One set of articles for your help centre and your web chat

The articles you publish in Web HQ are what your web chat AI answers visitors from. How that works in NearSync, what happens on its own, and what does not.

Karol Pelc26 October 20267 min read
Building NearSync
One set of articles for your help centre and your web chat
26 October 2026

An older post on this blog described NearSync as a layer for keeping content in step across other software. That was wrong. NearSync AI is one system for running a business: sales, WhatsApp and web chat, finance, HR and payroll, operations, projects and your website, with AI working in each of them.

There is one part of that old post worth keeping, though, because it points at a real problem. Most companies write the same answer twice. Once as a help article, and again as the script, macro or chatbot flow that answers the same question in a chat window. The two drift apart, and the day they disagree is the day a customer is told the wrong thing.

This post explains how we avoided that in NearSync, precisely: what happens on its own, what needs a person, and the decisions that cost us something.

Where the articles live

Your website tools live in Web HQ. For most businesses that means:

  • Knowledge Base, where you write, edit and publish articles, grouped into categories
  • Site Structure, for the pages of your site
  • Legal Manager, for terms, privacy and the other documents a site needs
  • Forms, Reputation for reviews, and a Web Overview of how the site is doing

The web chat itself is set up in Operations, under Web Chat, and the conversations arrive in the Support inbox next to your email and WhatsApp. So the chat is run by the team that answers customers, and the articles are written wherever your team writes them. What connects the two is the article.

Two switches decide who can read an article

Every article in the Knowledge Base has two switches.

  1. Published or draft. A draft is only ever seen by your team.
  2. Customer-facing. This one is off by default. Turn it on and the editor tells you what it means: the web chat can answer customers from this article.

An article reaches customers only when both are on. A published article that is not customer-facing stays internal. It is still useful, because your team can read it, but no visitor sees it and the chat will never quote it.

There is also an internal article type that cannot be made customer-facing at all. The switch is disabled in the editor, and if anything tries to save one with the switch on, it is turned off before it is saved.

We made customer-facing an opt-in on purpose, and it has a cost. The most common support question about the Knowledge Base is "I published it and it is not there". The answer is almost always that the second switch is still off. We could make that go away by turning the switch on by default. We will not, because the failure in the other direction is an internal pricing note or a draft refund policy read out to a customer by an AI. An article that is missing from the chat is an annoyance. An article that should never have been in it is a real problem.

What happens when you save

When you save an article, NearSync prepares it for the chat straight away. You do not press a separate button or wait for a nightly job.

  • The article is split into passages of about 2,000 characters, with about 200 characters of overlap so a sentence is never cut away from the one before it.
  • Each passage is turned into a form the AI can search by meaning, not just by matching words.
  • The new passages replace the old ones for that article. Edit a price in an article and the chat stops quoting the old price.

If you switch customer-facing off, unpublish the article or delete it, saving removes its passages from what the chat can see. Taking something back is the same action as putting it out, which matters when the thing you are taking back is wrong.

Behind that, a background check runs every 15 minutes. It looks for any article that should be available to the chat but is not, prepares up to 30 of them per run, and clears away passages from articles that no longer qualify. The save is how articles normally arrive. The check exists so that a save that failed quietly does not leave a hole that nobody notices.

A decision that came from an incident

That 15-minute check used to run every 3 minutes. In August 2026 the service we use to prepare articles hit its usage limit, and the frequent check kept retrying articles that could not be prepared, tens of thousands of times a day. Worse, the old order of operations removed an article's existing passages before preparing new ones. When preparation failed, articles that had been working became empty.

We changed three things:

  • New passages first, then the swap. An article's existing passages are only replaced once the new ones are ready. If preparation fails completely, the old passages stay and the failure is reported.
  • A budget and a cut-off. Each run prepares at most 30 articles and stops after two articles fail in a row, so an outage elsewhere costs almost nothing.
  • Every 15 minutes, not every 3. The check is a safety net, not the main path, so it does not need to be fast.

The cost is that an article whose save failed can take up to a quarter of an hour to reach the chat. We think that is the right trade. A chat that keeps answering from yesterday's article is better than one that answers from nothing.

How the chat uses an article

A visitor types a question into your web chat. Before any AI is involved, the message is checked against your chat intents. New organisations start with six:

  • Greeting, thanks and farewell get a short written reply, with no AI call.
  • Book a demo shows a booking card that links to your public meeting page.
  • Support, for messages like "something is broken" or "this isn't working", hands the chat to a person. The conversation is marked for your team and a support ticket is opened, so the request is tracked rather than lost in a chat log.
  • Pricing goes to the AI, with your articles.

Anything that does not match a pattern is classified by the AI against the same list. Questions that end up with the AI, including pricing, follow the same path:

  1. The question is matched by meaning against the passages from your customer-facing articles.
  2. The five closest passages above a similarity cut-off are given to the AI.
  3. The AI is told to answer in two to four sentences, to stay within those passages, to cite them, and not to invent details they do not contain.
  4. If nothing relevant is found, it is told to say so plainly in one sentence and offer to connect the visitor with a person.

The visitor can mark each AI reply as helpful or not. When a person on your team claims the chat from the inbox, the visitor is told someone has joined and the AI stops replying. A single chat also has a ceiling of 60 AI replies. After that, the AI hands over to your team and says so, rather than going quiet.

The AI that writes the reply is the model your organisation has chosen for it, running on NearSync's managed AI, so there are no keys to set up.

A bug the article editor exposed

One honest detail. Chat intents can narrow the search to articles tagged for that intent. The seeded pricing intent did exactly that: it only looked at articles tagged "pricing". But the article editor has no way to set that tag. So for every new organisation, pricing questions were matched against nothing, and the AI answered them with no articles at all.

We found it on our own site in October 2026 and fixed it on 24 October. The pricing intent now searches every customer-facing article. And whenever a narrowed search finds nothing, the chat searches everything once more before answering, so an article you wrote is not ignored because of a tag you could not set.

The help centre runs on the same articles

The public view of a knowledge base follows the same rule as the chat. An article can be read publicly only when it is published, customer-facing, and in a category that is public and not archived. That rule is enforced underneath the software, not in the page that shows the articles, so a mistake in a page cannot publish something private.

Our own help centre at nearsync.ai/help is built exactly this way, from the NearSync Knowledge Base. At the last build it held 306 articles in 28 categories, and they are the same articles the chat on this site answers from. When we correct an article, the help centre is corrected at the next build and the chat as soon as the article is saved.

If you are writing your own, our article on how to create a public knowledge base covers the practical side: start from your five most common support questions, title each article as the question somebody would type, and fix it the same day the thing it describes changes.

What this means in practice

  • Write the answer once. The article your customer reads is the article your chat answers from.
  • Correct it once. An edit changes what the chat says from the next question onwards.
  • Keep internal notes internal. Customer-facing is off until you turn it on, article by article.
  • Let the chat hand over. Support requests, unanswered questions and long chats go to your team in the Support inbox, with a ticket where it is a support request.

Web chat comes with every plan. Having the AI answer visitors in web chat and WhatsApp is an optional add-on, an AI agent seat, priced per agent on the pricing page.

To see how the pieces fit, read about web chat and the wiki and knowledge base, or book a demo and ask the chat on this site a question first. It is answering from the same articles this post describes.

See the product

Every module in one workspace. What shipped recently is in the changelog.

See the productRead the changelog

What changed for UAE businesses this month. We email once a month.