Here is a small mistake that can cost a shop a week of angry chats. Say a gadget shop raises its outside-Dhaka delivery charge from ৳100 to ৳130. Someone updates the FAQ but forgets the old "Shipping info" PDF sitting in the chatbot, and the bot keeps quoting the PDF.
The AI would be doing exactly what it was built to do: answering from the documents it was given. The problem was the documents. This guide covers what belongs in a chatbot's knowledge base, what does not, and how to write it so the answers come out right.
What is a chatbot knowledge base, and why does it decide answer quality?
A chatbot knowledge base is the content an AI chatbot searches before it replies: FAQs, policies, files, web pages and product data. It decides answer quality because a well-built chatbot answers only from what it finds there.
Each time a customer writes, the chatbot looks for the passages that best match the question and writes its reply from them. This is called retrieval-augmented generation. It means the AI is only as current, complete and consistent as your content. Feed it two conflicting delivery charges and it may pick either one.
What should I put in my chatbot's knowledge base?
Put in the facts customers ask about every day, written once each, plus the documents that answer longer questions. For a typical Bangladeshi online shop, that is this list:
| Content | Why it matters | Best format |
|---|---|---|
| Delivery areas, charges and times | The most asked question after price | Short Q&A pairs |
| Payment methods | "COD hobe?", "advance lagbe?" | Short Q&A pairs |
| Exchange and return rules | Prevents arguments after delivery | Policy note |
| Size guide and fit notes | "Amar height 5'4, kon size?" | Table in a text file |
| Product list | Price, options, stock, photos | Synced catalog or product entries |
| Shop details | Address, pickup, hours, phone | Short policy note |
| Care and usage | "Wash korle rong uthbe?", "warranty ache?" | Text file per category |
| Order process | "Order dile kobe confirm korben?" | Short Q&A pair |
That is the core. A shop with 30 well-written answers covers most of what customers ask on a normal day. Add more as you find gaps, not all at once.
What should I keep out of the knowledge base?
Keep out anything that is old, internal, unclear or duplicated, because the chatbot cannot tell which version you meant. The common culprits:
- Expired offers. An Eid offer left in the knowledge base keeps getting quoted in December.
- Drafts and internal notes. "Maybe free delivery over ৳3,000?" becomes a promise.
- Supplier price lists. Your cost price is not your selling price, and you do not want the bot to know either.
- Whole websites. Blog posts and campaign pages add noise. Import only the pages that answer customer questions.
- Scanned documents and photos of text. A scanned PDF holds an image, not text, so there is nothing for the chatbot to search.
- Legal boilerplate. A 4,000-word terms page buries the two sentences customers need. Write those two sentences as a Q&A instead.
The rule that prevents most mistakes: each fact lives in exactly one place. If the delivery charge is in an FAQ, it should not also be in a PDF, a web page and a product description.
How should I write knowledge base entries so the AI finds them?
Write each entry as a short, self-contained answer that uses the words customers use. The AI matches the customer's question against your content, so the closer your wording is to theirs, the better the match.
Compare two ways of writing the same fact:
Weak: "Logistics fees vary by destination zone per our courier partner's tariff."
Strong: "Delivery charge: Inside Dhaka ৳60, outside Dhaka ৳120. Dhakar bhitore ৳60, Dhakar baire ৳120. Delivery usually takes 1-2 days inside Dhaka and 2-4 days outside."
The strong version names the thing ("delivery charge"), gives exact numbers, includes a Banglish phrasing, and answers the follow-up question about time. Four habits make entries like this:
- One topic per entry. Delivery in one, returns in another. Long mixed documents get cut into pieces that lose context.
- The answer in the first sentence. Details after.
- Exact numbers and conditions. "Exchange within 3 days if the tag is attached" beats "easy exchange".
- The customer's words. If customers say "parcel", write "parcel", not "consignment".
Writing in Banglish as well as Bangla or English helps when many customers type that way. You do not need three copies of every answer. Adding the common Banglish phrasing to the question line of an FAQ is often enough.
How do I handle questions the knowledge base cannot answer?
Handle them with a fallback message and a handover to a person, never with a guess. No knowledge base covers everything, and the gap is where trust is won or lost.
A good fallback is short and honest:
Ei bishoye amar kache thik tottho nei. Amader team er ekjon apnake ektu porei reply korben.
Then make sure someone actually replies. Every unanswered question is also a free lesson: if three customers ask "Chattogram e same-day delivery ache?" in one week, that answer belongs in the knowledge base. Our guide to Banglish chatbot replies covers how to word fallbacks and answers for romanised Bangla.
How do I keep the knowledge base current?
Keep it current with a simple schedule tied to the moments your facts change. Most wrong answers come from stale content, not from the AI.
| When | What to check |
|---|---|
| Every price or charge change | Update the one entry that holds it, the same day |
| Start of a sale or festival | Add the offer with its end date written in the text |
| End of a sale | Delete the offer entry that day |
| Every month | Read 20 random chats and look for wrong or missing answers |
| After a courier or policy change | Update delivery times, areas and return rules |
Put the end date inside offer text ("Offer valid till 30 November") so that even if someone forgets to delete it, the chatbot can say it has ended.
How does the SaleMag knowledge base work?
SaleMag's AI chatbot answers from a knowledge base you build from several sources: PDF, Word, Excel, CSV, TXT, Markdown and JSON files up to 20 MB each, single web page URLs, whole sitemaps of up to 200 pages, question-and-answer FAQs, plain text, and your product catalog. The product source re-indexes when products change.
The bot is instructed to state facts only from that knowledge, your instructions and the customer's orders, and never to invent prices, policies or links. If nothing relevant is found, it sends your fallback reply, or hands the chat to your team when "hand over when unsure" is on. You choose how many knowledge snippets it reads per reply, from 1 to 20. The test playground shows which sources each reply quoted, so you can see when an old document is the problem. Smart (semantic) search needs an OpenAI or Gemini key; with only an Anthropic key, the bot uses keyword search. For product data specifically, see our guide on training a chatbot on your product catalog.
Next steps
Open last week's chats and list the 30 questions customers asked most. Write one short answer for each, in your customers' words, and delete any old document that says something different. Then test every question and check the source behind each answer. To build and test a knowledge base on your own content, start free with SaleMag or see pricing.
Step by step
- 1
List the top questions
Read a week of chats and write down the 30 questions customers ask most.
- 2
Write one answer per fact
Write each answer once, as a short question-and-answer pair or policy note, in your customers' language.
- 3
Add supporting documents
Upload size guides, care instructions and other text-based files, and add your product list.
- 4
Remove conflicts
Delete old offers, drafts and duplicate documents so no two sources disagree.
- 5
Test and review
Ask the top questions in a test chat, check which source each answer came from, and fix gaps.
Frequently asked questions
What is an AI chatbot knowledge base?
An AI chatbot knowledge base is the collection of content a chatbot searches before replying, such as FAQs, policy documents, web pages and product data. The chatbot finds the parts that match the customer's question and writes its answer from them.
Should I upload my whole website to the chatbot?
No. Upload the pages that answer customer questions, such as delivery, returns, size guides and contact details. Blog posts, old campaign pages and legal text mostly add noise and can lead to wrong answers.
Can the chatbot read scanned PDFs or images of my price list?
Usually not. A scanned PDF or a photo of a price list contains a picture of text, not text, so there is nothing to index. Type the content or export it as a text-based file before uploading.
How often should I update the knowledge base?
Update the knowledge base whenever a fact changes, such as a delivery charge, a return rule or an offer, and review the whole thing at least once a month. Remove festival offers the day they end.
What should the chatbot do when the answer is not in the knowledge base?
The chatbot should say it does not have that information or hand the chat to a person. It should never guess. Set a clear fallback message and turn on human handover for unanswered questions.
Should the knowledge base be in Bangla or English?
Write the knowledge base in the language most of your customers use. A mostly Bangla-speaking customer base gets better matches from Bangla FAQs, and you can add Banglish phrasings to the questions.
How big can an AI chatbot knowledge base be?
Size matters less than quality, because the chatbot only reads the few passages that match each question. In SaleMag, files can be up to 20 MB each and a sitemap import takes up to 200 pages.
About the author
WhatsApp & F-commerce automation, Dhaka
The Dhaka team that builds SaleMag, the AI sales inbox for WhatsApp, Messenger and Instagram used by online sellers in Bangladesh.