A knowledge base chatbot is a chatbot that answers customer questions from your own help centre articles, FAQs and documents rather than a fixed script. Modern ones retrieve the most relevant passages and have a language model write an answer from them, citing the source. They handle policy questions well, but cannot read a live order or take an action.
This guide is for an e-commerce operator weighing one up. It sits under our guide to AI agents and covers how the bot works, its types, what it cannot answer, set-up, risks and how to test one.
What is a knowledge base chatbot?
A knowledge base chatbot is a chat interface on top of your written help content. Your knowledge base is the record of how your store works: shipping, the returns window, sizing, care, warranty and payment options. A shopper asks “can I return sale items?” and gets an answer instead of a list of links.
You will also see it called a chatbot knowledge base (the content the bot answers from) or an AI knowledge base chatbot (the generative kind). It is the same category.
Our view is plain: the bot is only as good as the articles behind it. For the software that writes and maintains the articles, see our guide to knowledge bases.
How does a chatbot knowledge base work?
A knowledge base chatbot turns your articles into searchable passages, finds the passages closest in meaning to the customer’s question, then hands them to a language model with the instruction to answer only from them and say where the answer came from. The answer is only as current as the article.
The technique is retrieval-augmented generation. Lewis et al. introduced it in 2020 as a way to combine “pre-trained parametric and non-parametric memory for language generation”. AWS notes that it lets the model present answers with source attribution, and that a model’s training data alone is static with a cut-off date.
- Index. Articles are split into passages and converted to numerical representations (embeddings).
- Retrieve. The question is converted the same way and matched to the closest passages.
- Generate. The model writes an answer from those passages.
- Cite. It links or names the article it used.
Passage structure matters. Anthropic research shows that splitting documents into chunks strips context, so the system can miss the right passage. Contextual embeddings and contextual BM25 (keyword matching) together cut failed retrievals by 49%, and by 67% with reranking. Exact terms such as an error code still need keyword matching.
Citing the source is the honest part of the design: you and the customer can click through and check. But a citation shows which article was used, not why the bot concluded what it did. We come back to that under risks.
What types of knowledge base chatbot are there?
Three types of knowledge base chatbot exist, and the difference is what each answers from and whether it can do anything. A keyword FAQ bot matches words to a stored answer. A retrieval-augmented bot writes an answer from your articles. An AI agent also reads live store data and takes the action, under your rules.
The examples below are illustrations with a fictional 30-day returns window, not a real store’s policy.
| Type | How it finds the answer | Answers from | ”Can I still return these trainers?" | "Where is my order?” | Can it act? | Typical failure |
|---|---|---|---|---|---|---|
| Keyword FAQ bot | Keyword or decision-tree match to a preset answer | A fixed list of question and answer pairs | Pastes the generic returns policy, whatever the order | Shows a “track your order” link | No | Breaks on phrasing it was not given, then loops |
| Retrieval-augmented bot (an AI knowledge base chatbot) | Meaning-based retrieval plus a language model | Your help centre, FAQs and documents | Explains the 30-day window in plain words and cites the article, but does not know when this order was delivered | Explains how tracking works, cannot see this parcel | No | Confident wrong answers when articles conflict or are out of date |
| AI agent with live data and actions | Retrieval plus live order, catalogue and customer data, plus tools | Articles, policies and live store data | Checks the delivery date on the order, applies the policy, starts the return if it qualifies | Reads the order and carrier status and gives the actual answer | Yes, within the rules you set | A wrong action if the rules are loose |
The CFPB describes the first kind as using “decision tree logic or a database of keywords to trigger preset, limited responses”. It also names the result: “doom loops”, where an issue falls outside the bot’s limits and the customer has no way to a person. A FAQ chatbot of this kind is easy to outgrow.
For the full comparison, see our agent comparison.
What can a knowledge base chatbot answer, and what can’t it?
A knowledge base chatbot answers questions whose answer is the same for every customer and is written down: returns windows, shipping countries and costs, sizing, care, warranty and payment options. It cannot answer questions that depend on this customer’s order, or that need something done, such as where is my order, refund status, address changes and cancellations.
| Customer question | Knowledge base chatbot | What it needs to resolve (our view) |
|---|---|---|
| What is your returns window? | Answers well | A current returns article |
| Do you ship to Norway, and what does it cost? | Answers well | An up-to-date shipping article |
| Which size should I take in this fit? | Partly: a general size guide only | Catalogue data per product and fit |
| Is this in stock in M? | No: articles do not hold live stock | Live inventory |
| Where is my order? | No: explains tracking in general | The order, fulfilment and carrier status |
| Has my refund gone through? | No | The order and payment state |
| Can you change my delivery address? | No: can only explain the process | An action in the store, within your policy |
| My item arrived damaged | Explains the claims process | Order data, policy and an action (replacement or refund) |
The gap is big for a store. LateShipment estimates that somewhere between 20 and 40% of support tickets are people asking where their order is. A knowledge base chatbot cannot close any of those.
Our customers show where the line sits. At MR MARVIS the agent handled WISMO requests by checking backend data instantly. Otrium’s team had been busy with the same repetitive questions about returns, refunds and order status, which is the work behind refunds and returns. For sizing and fit, see product advice.
If you are choosing a bot for an online store, our guide to e-commerce chatbots compares them. For the wider picture of store bots, see store chatbots.
What are the benefits for an online store?
A knowledge base chatbot gives an online store fast answers to policy questions and takes repeat questions about returns windows, shipping and care off the support queue. It helps most where the answer is the same for every customer and is already written down in your help centre.
- Fast answers on policy. The shopper gets the rule in plain words instead of a list of links.
- Fewer repeat policy tickets. Your team spends less time retyping the same returns and shipping answers.
- A visible source. The customer and your team can click through to the article and check it.
How do you set up a knowledge base chatbot?
Set up a knowledge base chatbot by fixing the content first, then switching on the bot. Make the articles accurate and self-contained, decide what the bot says when it does not know, and make sure a person can take over with the full conversation.
Content is where most teams are weakest. A Gartner leaders survey found that 85% will explore or pilot customer-facing conversational GenAI in 2025, yet 61% have a backlog of articles to edit and more than one-third have no formal process for revising outdated ones. Senior Principal Kim Hedlin said leaders “cannot ignore existing issues with knowledge management”.
Articles. Give each one an owner and a review date, and put exceptions next to the rule. Writing and upkeep are covered in our guide to knowledge bases.
Fallbacks. Tell the bot to say it does not know rather than guess. Kalai et al. explain in 2025 that language models “sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty”.
Hand-off to a person. The bot should tell the customer it will connect them to a person, then let the chat continue where it left off. Pass the transcript, the order it concerns and the reason for the hand-off. Our guide to automated contact covers hand-offs in detail, and the automation guide has the wider how-to.
What are the risks of a knowledge base chatbot?
A knowledge base chatbot carries three risks. Stale articles give out-of-date answers with full confidence. Retrieval and generation can still produce a wrong answer, even with a citation. And most bots leave no record of why they answered the way they did, so you cannot find and fix the cause. You are accountable for every answer.
Stale and conflicting articles. They give out-of-date answers with full confidence. Keeping them current is covered in our guide to knowledge bases.
Wrong answers that cite the right page. In Moffatt v. Air Canada (tribunal ruling, 2024 BCCRT 149, 14 February 2024), the airline’s website chatbot told a customer he could apply for a bereavement fare retroactively. It even linked to the airline’s “Bereavement travel” page, which said the opposite.
The tribunal found Air Canada “did not take reasonable care to ensure its chatbot was accurate”, said “It makes no difference whether the information comes from a static page or a chatbot”, and awarded CAD 650.88 in damages. It is an airline, not a store, but the lesson carries: a citation next to an answer does not make the answer right.
Confident wrong answers, even with retrieval. Stanford RegLab and HAI tested retrieval-based legal research tools in May 2024. Leading tools “produced incorrect information more than 17% of the time”, and one more than 34%. Retrieval reduced errors against general-purpose models, but “even RAG systems are not hallucination-free”. They counted both wrong answers and answers citing a source that does not support the claim. It is a legal test, so read it as evidence about the method, not about support bots.
No audit of why it answered. When RAG was introduced, Lewis et al. listed “providing provenance for their decisions” as an open research problem. A source link tells you which article was retrieved, not how the bot reasoned from it or which rule it applied. Without that, you fix symptoms one article at a time.
Customer trust. A bot that makes it harder to reach a person costs you trust, which is why the hand-off matters.
| Risk | What it looks like in a store | How to reduce it |
|---|---|---|
| Stale article | A Black Friday returns window still quoted in January | An owner and a review date per article |
| Conflicting articles | Two shipping pages with different cut-off times | One article per rule |
| Wrong answer with a citation | The bot paraphrases a policy into something it does not say | An “I don’t know” fallback and a check of each answer against its source |
| No reasoning trail | You cannot see which rule it applied | A tool that shows its reasoning and source |
How do you evaluate a knowledge base chatbot?
A knowledge base chatbot should be tested on your own real questions before launch, not on a demo. Check whether answers match your current policy word for word, whether it says “I don’t know” when it should, whether it shows the source and its reasoning, and whether it hands over with context. Then check what share of questions need live order data.
Our suggested method:
- Run 50 to 100 recent real customer questions through it.
- Mark each answer correct, wrong or “should have handed over”.
- Check that the cited article says what the answer says.
- Count how many questions needed order data or an action. That is the share a knowledge base chatbot cannot close.
For vendor choice, our guide to the best agents is the place to start.
Where does Engaige fit?
Engaige (that’s us) isn’t a knowledge base chatbot: it’s the e-commerce-native AI support agent that plugs into the help desk you already run, answers from live store data, its policies and the docs you add, and takes the action. Engaige AI configures your support agent from your conversations and inputs, and shows which policies it used. Otrium has 60% of 120,000 annual tickets resolved end to end.
The difference is policies, not just articles. Engaige works from structured policies that set how it reasons, what process it follows and when it may take each action, so the same policy is applied the same way each time. You decide what it may do on its own and where a person checks first.
Scenario
The customer asks about a delivery address.
Instructions
Read the order with Get order , then take the branch that matches.
Get order Scenario
The order has not been picked yet.
Instructions
Move it with Change picqer shipping address .
Change picqer shipping address Scenario
The parcel has already left the warehouse.
Instructions
Say it cannot be redirected, then offer Process refund .
That answers the audit risk too: every answer and action shows why, with the source, and Engaige AI lets you test every policy before launch.
Engaige offered control, flexibility, and the ability to really incorporate AI in a more human way.
You do not need a finished knowledge base first: Engaige AI drafts your initial policies from the integrations you enable and a little information you give it, you go live in Agent Assist in hours, and it improves from your team’s replies under your approval.
Ticket types then go autonomous one at a time, most brands in about a week, up to 90% of supported tickets; the full loop is in our guide to AI agents, or see a demo.
Keep your helpdesk. Replace the AI.
Frequently asked questions
What is a knowledge base chatbot?
A chatbot that answers customer questions from your own help articles, FAQs and documents, using a language model to write the answer and cite the source.
How does a knowledge base chatbot find its answers?
Through retrieval-augmented generation. It finds the passages closest in meaning to the question, writes an answer from them and cites the article it used.
What is the difference between an FAQ chatbot and an AI one?
An FAQ chatbot matches keywords to preset answers. An AI knowledge base chatbot writes an answer from your articles.
Can a bot that answers from articles tell a customer where their order is?
No. That needs live order and carrier data. LateShipment estimates that where-is-my-order questions are 20 to 40% of support tickets.
Why do bots that answer from articles give wrong answers?
Stale or conflicting articles, retrieval misses and models that guess instead of saying they do not know. The Air Canada case shows a bot contradicting the page it cites; the Stanford study shows retrieval tools still get answers wrong.
How do you keep a chatbot knowledge base accurate?
Give each article an owner and a review date, keep one rule per article with its exceptions and make each section self-contained.
Is Engaige a knowledge base chatbot?
No. Engaige is the e-commerce-native AI support agent that works inside the help desk you already run. It answers from live store data, its policies and any docs you add, and takes the action.
Do I need a complete knowledge base before I use an AI support agent?
No. Engaige AI drafts your initial policies from your integrations and some initial information. You go live in Agent Assist in hours, and it improves from your team’s replies under your approval.
