Knowledge Base Chatbot: What It Is and What You Actually Need

If you search for "knowledge base chatbot" (a chatbot that answers from information about your business), most of what you find describes the same system. It takes in your documents, splits them into small sections, converts each section into a numerical form a computer can compare (an embedding), stores the results in a vector database, and pulls out the closest matches when somebody asks a question.

That system works, and it is the right choice if you have ten thousand support articles. If you run a dental practice, a law office, or a home services company, you almost certainly do not. You have a page of pricing, a page of policies, a list of what you do and do not do, and about forty questions customers ask you again and again.

For that amount of information, such a system is more than you need. This guide explains what a knowledge base chatbot is, the two ways to build one, how to tell which one your business needs, and what separates a chatbot that answers well from one that confidently makes things up.

What a knowledge base chatbot actually is

A knowledge base chatbot is a chatbot that answers from information you gave it, rather than from general training or a script you wrote in advance.

This is the key difference. A scripted chatbot follows a flow (conversation paths you build in advance) and can only handle those paths. If a visitor asks something you did not plan for, it replies "I did not understand that" or returns them to a menu. A knowledge base chatbot reads the question as asked, checks the information it has about your business, and answers in its own words.

The main benefit is that it can handle the many unusual questions that each come up only occasionally. You are unlikely to build a scripted path for "do you take my insurance if I am out of state" or "what happens if I need to reschedule twice." Each of these questions is rare on its own, but together they come up all the time, and they are often the reason a visitor leaves your site to call a competitor.

The two approaches compared

There are two ways to give a chatbot your information. Most articles describe only the first, because it is the approach that large software vendors usually sell.

Retrieval, also called RAG

With retrieval-augmented generation (RAG), you upload documents. The system splits them into sections, converts each section into an embedding, and stores the results. When a question arrives, it converts the question the same way, finds the closest sections, and passes them to a language model (the AI that writes the reply) to compose the answer.

Strengths: it can work with a collection of documents far too large to send to the AI all at once, and it can point to the source document for each answer.

Drawbacks that are rarely mentioned: you now have a system to maintain. When documents change, the index (the searchable copy of your documents) has to be rebuilt. A policy can be split across two sections, so the chatbot answers with only half of it. The search can also return the wrong sections, and the AI then writes a confident answer from the wrong source, which is worse than no answer because it sounds authoritative. Much of your time goes into fixing the search rather than writing content.

Prompt-loaded knowledge

With prompt-loaded knowledge, you write what the chatbot should know into a single field. In every conversation, that text is sent to the AI along with the visitor's question; the combined text is called the prompt. There is no index to rebuild and no search step that can go wrong.

Strengths: there is nothing to rebuild, changes take effect as soon as you save, and the AI sees all of your information at once instead of only the few sections a search selected. The last point is especially important. If a visitor asks "how much is a cleaning and do you take my insurance," a retrieval system has to find both answers separately, while an AI that has the whole document can answer both parts in one reply.

The limit: there is a maximum size. Everything you write is sent to the AI with every message, so the field cannot hold a full manual. This is a firm limit.

Which one you need

The best test is the amount of information your customers actually need, rather than how much you might like to include.

Alma limits its knowledge field to 50,000 characters, which is roughly 8,000 words or about sixteen pages of dense text. Before you decide that is too small, write down everything a customer needs to hear from you. That usually means your services and what each includes, your prices or the reason you do not publish them, your hours, location and parking, the work you do not do and who you refer it to, your cancellation policy, and the dozen questions you answer every week.

This article is part of our guide to chatbot knowledge and measurement.

For most small businesses, the finished list fills only a few pages, well under the sixteen-page limit.

Retrieval makes sense when you have a large library, such as hundreds of individual products (each with its own stock-keeping unit, or SKU) and distinct specifications, a decade of technical documentation, or regulatory material that has to be quoted word for word. If that describes your organization, choose a system built for that purpose.

The chance that your information might grow later is not a reason to choose retrieval. If your information fits, prompt-loaded knowledge is faster to set up, cheaper to run, and easier to keep accurate than a retrieval system nobody has time to maintain.

What good knowledge looks like

The quality of what you write decides whether your chatbot is useful. Many vendors say little about it, because it is work you have to do rather than a feature they can sell.

Write plain answers rather than marketing language. A line such as "We pride ourselves on flexible scheduling" gives the chatbot nothing useful to say. A line such as "Appointments can be rescheduled up to 24 hours ahead at no charge. Inside 24 hours there is a $35 fee" gives it a clear answer.

State your limits clearly. Some of the most useful lines describe what you do not do. A line such as "We do not service commercial properties" prevents the chatbot from wrongly saying yes, and it saves you from taking calls about work you do not offer.

Use the words your customers use. If your team says "intake" and customers say "first appointment," write both. The AI can recognize words with similar meanings, but it cannot know a term you never wrote down.

Say what to do when it does not know. Tell it to offer the contact form or the phone number rather than guess. A chatbot that says "I am not sure, here is how to reach the team" is working correctly. A chatbot that invents an answer to sound helpful can mislead your customers.

Keep it current. Outdated information is a common problem in practice. When prices change but the knowledge field does not, the chatbot keeps quoting the old prices. Schedule a regular review on your calendar. With prompt-loaded knowledge, an update means editing one field, which makes this approach practical for a busy owner.

It is also worth choosing the chatbot's tone deliberately, because it will have a tone whether you choose one or not. That topic is covered in chatbot personality and tone.

What it will not do

There are two limits to keep in mind, because a chatbot that promises too much creates the support requests it was meant to prevent.

It does not know anything you did not tell it. Alma does not read your website or import your files. If your hours are not in the field, the chatbot does not know them.

It is not a replacement for a human on hard conversations. Complaints, refunds, and anything involving money or strong feelings should reach a person quickly. The chatbot's job is to handle routine questions so that your staff have time for the ones that are not routine.

Setting one up

Alma sets this up as a done-for-you build rather than a settings screen, and that choice is deliberate. The difficult part of a knowledge base chatbot is not the software but writing the answers, and many self-serve tools leave that step to you, which is often where people give up.

1. You fill in an intake form. It asks about your services, hours, pricing or why you do not publish it, where leads (interested visitors' contact details) should go, and the questions customers actually ask. This gathers your knowledge through questions instead of a blank field.
2. We build the bot. Scripted flows handle your main tasks, such as booking, quoting and collecting a lead's contact details, because those paths should work the same way every time. Your answers become the knowledge the chatbot draws on for everything else.
3. You review it before it goes live. This step deserves careful attention. Ask it your most difficult questions rather than the easiest ones, including unusual cases, topics you would rather not discuss, and questions you find tiresome. Each weak answer shows us a line to add.
4. It goes on your site. It is added with one short piece of code (a snippet), in the same way as any other chat window, or widget.
5. We keep it current. Prices and policies change, and outdated knowledge is the usual reason these chatbots stop answering well. Because the knowledge is one document rather than an index, an edit takes effect immediately, with nothing to rebuild.

The build is a one-time white glove setup, meaning we do the setup work for you. AI answering then runs as a $25 per month add-on for 500 AI-answered messages, quoted separately so that a recurring charge is never a surprise. Flows work without it, so a bot that only books appointments costs you nothing extra. Current rates are on the pricing page.

Where it fits with everything else

Answering from a knowledge base is one of several tasks the same chatbot can do.

If your goal is to reduce repeated questions, the related article on automating FAQ responses explains which questions are worth automating and which are not. If you want customers to resolve issues completely without contacting anyone, read about customer self-service. If you are replacing a queue of support requests rather than a phone line, see modernizing the help desk.

Two other articles are worth reading before you go live. Your knowledge field will contain business information, so the article on security and privacy is worth ten minutes. If the leads your chatbot collects need to reach the software where your team keeps contact records, see CRM integration.

Common questions

Does it get trained on my documentation?

No. The AI is not trained or adjusted on your information. Instead, your knowledge is sent to the AI each time a question is asked. As a result, an edit takes effect immediately, without any retraining or rebuilding of an index.

Can I upload a PDF or point it at my website?

Not in Alma. Your knowledge is text that you write or paste into the field. If your information is in a PDF today, the task is to condense it into the answers customers need, which is worth doing whichever tool you use.

What happens when it does not know?

If you have told it what to do, it says that it does not know and tells the visitor how to reach your team. That instruction belongs in your knowledge field, and it is one of the most valuable lines you can write.

Will it make things up?

The risk is real, but it can be managed. Two things greatly reduce it: knowledge that states your limits clearly, and an instruction to refer the visitor to your team rather than guess. When it does happen, the cause is usually a gap in what you wrote rather than the AI inventing details on its own.

How much knowledge is too much?

Alma's field holds 50,000 characters, roughly sixteen pages. If your text is longer than that, the usual cause is pasting in source material rather than writing answers. A twelve-page policy document can usually be condensed to one page covering what customers actually ask about it.

Does this replace my scripted flows?

No. Flows handle the important paths, where you want the same questions in the same order every time. AI answering handles everything else. Most chatbots that work well use both.

Summary

A knowledge base chatbot answers from the information you give it. If your business information fits in about sixteen pages, as it does for most small businesses, you do not need a document-processing system or a vector database. You need a well-written page of answers, an instruction about what to do when the chatbot does not know, and a habit of reading the transcripts.

Writing the knowledge takes more effort than setting up the technology, and it is what decides whether the chatbot is useful.

Start a free trial and build the flow first. You can add AI answering later to handle the questions you did not plan for.

To start from an example, see templates with a knowledge base already wired.