In short: A chatbot knowledge base is the set of sources an AI assistant is allowed to answer from. In Puhera, you build it by letting the platform crawl your website, uploading PDF, Word, PowerPoint and text files, adding an FAQ and syncing your product catalog. For accurate answers, keep each document focused on one topic, keep content current and test the bot with real customer questions.
What is a chatbot knowledge base, and why does it matter so much?
The knowledge base works like your bot's memory. When a visitor asks a question, the AI first looks for relevant information in these sources, then writes an answer grounded in what it found. This approach is usually called RAG (retrieval-augmented generation): retrieve the right passage, then generate the reply from it.
The consequence is simple: a bot is only as good as its knowledge base. Missing, contradictory or outdated content leads to weak answers no matter how capable the underlying model is. That is why a knowledge base should be treated as a living content asset, not a one-time setup task.
In Puhera, the bot only speaks from your own information and shows the source behind every answer. That makes it easy to see which document produced which reply and to trace a wrong answer back to its cause. We explain why that matters in our post on AI chatbot answers with citations.
Which sources should you add to your knowledge base?
Every business is different, but on most sites four types of sources do the heavy lifting. Start with the sources that cover the questions customers ask most often, then expand.
- Website crawl: Makes your public content readable to the bot, such as your about, shipping, returns, payment, services and contact pages. If your site is already well maintained, this is the fastest place to start.
- Documents (PDF, Word, PowerPoint, text): Ideal for information that does not live on your site, such as user manuals, technical specifications, service brochures and process notes.
- FAQ / knowledge base entries: Short, clear answers to your most common questions keep the bot consistent where consistency matters most.
- Product catalog: If you sell online, syncing product data from a Google Merchant feed or from T-Soft lets the bot answer product questions from catalog data rather than guesswork.
For the catalog side in detail, see our guide to using a Google Merchant feed as a chatbot product catalog.
How should you structure documents so answers are accurate?
AI performs best when the information it needs is easy to find and stated plainly. The same fact written five different ways in five different places creates confusion. Run your documents through this checklist before uploading them:
- One document, one topic. Keep your return policy, shipping terms and warranty conditions in separate files instead of one giant document.
- Use descriptive headings. Write "International shipping times" rather than "Section 3" so each part says what it covers.
- Lead with the answer. The first sentence of each section should state the key fact; explanations come after.
- Remove contradictions. If an old brochure says 14 days and your site says 30 days for returns, the bot cannot know which one is right. Keep a single source of truth.
- Spell out abbreviations and internal jargon. Customers do not know your team's shorthand. Write the full term the first time it appears.
- State conditions and exceptions. If "free shipping" depends on a minimum order or a region, say so in the same sentence.
Picture a clothing store that keeps its size chart, fabric care instructions and exchange terms mixed together in one PDF. Asked "Can I exchange these trousers?", the bot may reach for a sentence about washing temperatures. Split the same content into three well-titled documents and the right passage becomes easy to find, and the cited source matches the question.
What should you leave out?
More content does not automatically mean better answers. Irrelevant or sensitive material can lead the bot to the wrong context, or worse, to sharing something it should not.
- Drafts, old versions and files marked as superseded
- Documents containing personal data about customers or employees
- Internal pricing negotiations, contract details or anything meant only for your team
- Promotion pages for campaigns that have already ended
A simple rule: if you would not be comfortable publishing a piece of information on your public website, do not put it in the knowledge base of a customer-facing bot.
How do you keep a chatbot knowledge base fresh?
The most common mistake is to build the knowledge base and forget about it. As prices, delivery times and campaigns change, the bot has to change with them. Otherwise it will confidently share outdated information, complete with a source.
- Assign ownership. Give each type of source an owner: the support team owns the return policy, the e-commerce team owns product data.
- Update at the moment of change. When a policy changes, replace the old file with the new one. Never leave both in place.
- Sync your catalog instead of retyping it. Catalog sync makes it far easier for product changes to reach the bot than maintaining product facts by hand.
- Review on a schedule. Once a month, look at the questions the bot receives most and the documents it cites for them.
How do you test a knowledge base with real questions?
Before going live, test the bot with the questions your customers actually ask. You can collect them from support e-mails, phone notes and direct messages.
- Build a question list. Write down your 20 to 30 most frequent questions in the customer's own words, typos and casual phrasing included.
- Ask them in preview. For each answer, note whether it is correct and which source it cites.
- Fix wrong sources. If the bot pulls the right fact from the wrong document, simplify the document or remove the conflicting content.
- Fill the gaps. Every question the bot cannot answer points to missing content. Add it.
- Probe the edges. Ask off-topic questions and questions the knowledge base does not cover, and check that the bot handles them honestly instead of improvising.
Testing does not stop at launch. Real conversations are the best signal you will get about which topics are still missing.
How does Puhera help you build a knowledge base?
Puhera's AI knowledge base brings all of the sources above together in one place, without code:
- It crawls your website automatically.
- It accepts PDF, Word, PowerPoint and text file uploads.
- It links your FAQ / knowledge base content.
- It syncs your product catalog from a Google Merchant feed or T-Soft.
- It shows the source behind every answer, so during testing you can see exactly which document produced what.
On the Pro plan you can also connect tools such as Notion through MCP (Model Context Protocol, an open standard for connecting external data sources to AI models); our post on what MCP is covers the details. Once your sources are in and you have tested in preview, you publish the bot with a public link or an embed script. You can start a 7-day free trial with no credit card.
Frequently asked questions
Which file types can I use for a chatbot knowledge base?
In Puhera you can upload PDF, Word, PowerPoint and text files and have your website crawled automatically. FAQ content and a product catalog synced from Google Merchant or T-Soft can also be part of the knowledge base.
How often should I update my knowledge base?
Update it as soon as a price, policy or product detail changes. Beyond that, reviewing the bot's most frequent questions and the sources it cites once a month is a good habit.
What should a bot do when the answer is not in the knowledge base?
A well-designed bot should say it does not have the answer instead of making one up, and point the visitor to the right channel. In Puhera you can set up an automation that connects the visitor to a live agent in those situations.