How to Build a Chatbot Knowledge Base Your AI Agent Can Answer From

A chatbot knowledge base is the set of approved content your AI agent searches before it replies: answers to common questions, product and service details, policies, locations and opening hours. When a customer asks something, the agent retrieves the most relevant passages and writes its answer from them. Build it from the questions customers really ask, write each fact once in plain language, split it into focused sections, and update it whenever your business changes.

Most disappointing AI agents are not limited by the model. They are limited by what they can look up. If your return policy lives in a scanned PDF from three years ago, the agent will either quote it or guess. This guide covers what to include, how to structure it, how retrieval works in plain words, and how to keep the content accurate once customers are talking to it.

What a chatbot knowledge base actually does

A large language model knows a lot about the world and nothing about your business. It does not know your prices, your branches or whether you deliver on Sundays. A knowledge base closes that gap.

The standard technique is retrieval-augmented generation (RAG). In plain words:

1. The customer asks a question on WhatsApp, Instagram or your website chat.
2. The system searches your knowledge base for the passages that best match the meaning of the question, not just its exact words.
3. The agent receives those passages together with the question and its instructions.
4. The agent writes the reply from the passages, ideally saying nothing the passages do not support.

This is why you do not need to “train” or fine-tune a model to make it answer from your documents. You change the content, and the next answer changes with it.

What to include in your knowledge base

Start from customer questions, not from the documents you happen to have. Pull the last few hundred conversations from your inbox, your email support and your phone notes, and group them by topic. The topics that come up most are your first sections.

Content type	Examples	Why the agent needs it
FAQs	”Do you ship abroad?”, “Can I change my booking?”	Covers the bulk of everyday questions
Products and services	Names, variants, what is included, who it suits	Lets the agent recommend and compare
Prices and fees	Price lists, delivery fees, deposits	Prevents invented or outdated prices
Policies	Returns, cancellations, warranty, privacy	High-risk answers that must be exact
Locations and hours	Addresses, opening hours, holiday closures	Simple facts customers ask constantly
Processes	How to book, how to pay, how to claim warranty	Turns answers into next steps
Escalation rules	What the agent must never answer itself	Tells it when to hand over

Leave out internal notes, drafts, old campaign terms and anything with personal data. A knowledge base is something the agent may quote to any customer, so treat every line as public.

How to structure content so retrieval finds it

Retrieval works on pieces of text, not whole files. How you write and split content matters more than how much you have.

Write one topic per section

A section titled “Returns for online orders” with five short paragraphs retrieves far better than a 40-page terms document where returns appear on page 31. If a page covers several topics, split it.

Put the answer first

Start each section with the direct answer, then add conditions and exceptions. The passage that gets retrieved should contain the answer, not only the background.

Weak:  Our company was founded in 2009 and has always cared about
       customer satisfaction. Over the years our returns process...

Better: Online orders can be returned within 14 days of delivery.
        Items must be unused and in their original packaging.
        Sale items can be exchanged but not refunded.

Write each fact once

If your delivery fee appears in four documents and you update only three, the agent may retrieve the old one. Keep one source of truth per fact and remove duplicates.

Use the words customers use

Customers write “money back”, not “reimbursement procedure”. Add the phrasings you see in real chats as question titles, so the meaning search has more to match.

Keep tables and lists simple

Simple price tables and bullet lists usually survive splitting well. Complex merged-cell spreadsheets and scanned images often do not. When in doubt, rewrite them as short sentences.

Chunking in plain words

Before content can be searched by meaning, it is cut into chunks, and each chunk is converted into a list of numbers called an embedding, which is stored in a vector database. At answer time, the question is converted the same way and the closest chunks are returned.

Two settings shape the results:

• Chunk size. Small chunks are precise but can lose context (“within 14 days” without saying of what). Large chunks keep context but bring in unrelated text, which can confuse the answer.
• Overlap. Repeating a little text between neighboring chunks reduces the chance that a key sentence is split in half.

You rarely need to tune these on day one. Sensible defaults work for well-structured content. Revisit them only when tests show the agent retrieving the right document but the wrong part of it.

Separate knowledge by purpose

One shared pile of documents makes answers worse as it grows. A sales assistant does not need your internal shift procedures, and a support agent should not quote a draft price list. Split content into separate domains, such as customer support, sales and internal operations, and give each agent only the domains it needs.

This also reduces risk. Content that an agent cannot retrieve is content it cannot leak. For more on what data should and should not reach an AI system, see our AI chatbot privacy checklist.

Combine static knowledge with live data

A knowledge base holds content that changes slowly: policies, product descriptions, opening hours. It is the wrong place for an order status, stock level or a customer’s next appointment. Those should come from your systems at the moment of the question, through tool calls (function calling) or integrations.

A good rule: if a fact can change several times a day, fetch it live. If it changes a few times a year, put it in the knowledge base.

Keep the knowledge base current

A knowledge base that was right at launch drifts quickly. Build maintenance into your routine:

1. Assign an owner for each section, for example operations owns hours and locations, finance owns prices.
2. Update on change. When a policy or price changes, update the knowledge base the same day, then re-index it.
3. Review unanswered questions weekly. Every question the agent could not answer is a gap, a wording problem or a case for a human.
4. Run a monthly audit. Read the top sections, remove outdated content and check that seasonal information (holiday hours, campaign terms) has expired.
5. Watch handovers. If the same topic keeps going to people, either add content or decide it should always be handled by a person. Our guide to AI-to-human handoff covers the rules.

How to test a knowledge base before launch

Write a test set of 30 to 50 real questions and the answer you expect for each. Include:

• Common questions phrased the way customers write them, including typos and short messages.
• Edge cases where the policy has exceptions.
• Questions that are not covered, to confirm the agent says it does not know rather than guessing.
• Questions in other languages if you serve international customers.

Test: "can i send back shoes i wore once"
Expected: Explains that returns must be unused, offers to connect
          the customer with the team for exceptions.
Fail if: Agrees to a refund, or invents a different time limit.

Check two things separately: did the system retrieve the right passage, and did the agent use it correctly? A wrong answer from the right passage is a prompt problem. A wrong answer from the wrong passage is a content or chunking problem. Our guide to testing an AI agent before launch goes further into test sets and metrics.

Common knowledge base mistakes

• Uploading every document you own instead of the content customers need.
• Keeping prices or delivery promises in the agent’s instructions, where nobody remembers to update them.
• Leaving old versions of policies next to new ones.
• Treating scanned images or complex spreadsheets as readable content.
• No fallback when nothing relevant is found, which invites hallucination.

How to build your AI knowledge base in MonoChat

MonoChat gives you two ways to build what your agent knows, depending on how much control you want.

The quick start: MonoChat Start

MonoChat Start reads your website and writes the knowledge base for you: a business profile, your products and services, your locations, your policies and around forty questions your customers actually ask. No website, or one with little to read? You answer five questions instead.

Everything the assistant may say sits on four tabs: Questions, Products & services, Locations and Policies. You can edit any entry, switch it on or off, or export it. The assistant never states a price, delivery time or guarantee that is not written there. Answers it wrote from general knowledge about your sector, rather than from your site, arrive switched off with a “Check this” badge until you approve them.

When it has no answer, it does not improvise. It writes the request down under Messages taken and tells the customer a person will come back to them, and its reports list the questions it could not answer, which is your weekly update list. It answers on WhatsApp, Instagram, Telegram or your website chat.

Full control: the AI knowledge base and RAG

For larger content sets, the AI knowledge base lets you upload PDFs, Excel files, text and web pages, with bulk upload, metadata and tags. You organize content into scopes, such as customer support, sales or internal procedures, each with its own access control, and assign each scope to the agents and flows that should use it.

MonoChat chunks and embeds the content with ready-to-use defaults, and you can choose the embedding model, chunk size, overlap and preprocessing rules when you need to. Vectors are stored in MonoChat’s embedded vector databases, or in your own through VectorDB integration. When content changes, you update the scope and re-index it, and the assistant retrieves from the new version without you rebuilding your flows.

To mix documents with live data, RAG orchestration chains steps such as query rewrite, retrieval, reranking, tool calls, citations and the final response, and AI function tools fetch things like an order status or a CRM record in the same answer.

You can start on the free Basic plan (one agent, one channel) at app.monochat.ai and compare plans on our pricing page.

Sources

• Meta for Developers: WhatsApp Business Platform overview
• WhatsApp Business Messaging Policy