The difference between a chatbot customers love and one they close immediately is not the AI model — it is what the AI knows. A bot that answers "our delivery fee is Rs. 150 inside Ring Road, free above Rs. 2,000" earns trust. A bot that says "delivery fees may vary" wastes everyone's time.
This is a practical guide to teaching an AI assistant your business, without writing code.
Where the knowledge comes from
Modern no-code platforms build a knowledge base from three sources:
- Your website. Point the platform at your domain and it reads your pages — services, prices, policies, FAQs — and converts them into structured knowledge.
- Documents. Menus, price lists, brochures and policy PDFs get uploaded and extracted the same way.
- Manual entries. Quick question-and-answer pairs you type in for anything not written down anywhere — "do you take card payments at the counter?" is often tribal knowledge.
In Sakhi.Chat all three feed one knowledge base per business, and every channel — WhatsApp, Instagram, web chat, email and the rest — answers from it.
How the AI uses it (and why bots hallucinate)
When a customer asks a question, the platform retrieves the most relevant pieces of your knowledge and instructs the model to answer only from that context. Hallucination — the bot confidently inventing an answer — happens when that instruction is missing or weak, or when the knowledge base simply does not contain the answer.
Two design choices keep a bot honest:
- Grounded answering. The bot is told to answer from your content and say so when it cannot, rather than improvising.
- Human handover as the escape hatch. "Let me connect you with the team" is the correct answer to a question outside the knowledge base.
Finding the gaps before customers do
The fastest way to improve a bot is to look at what it failed to answer. Sakhi.Chat's knowledge-gap detection flags questions the AI could not answer confidently and clusters them, so every week you can turn real customer questions into new knowledge entries. Ten minutes of gap review a week compounds quickly.
A 30-minute setup that actually works
- Connect your website and let it scrape your public pages.
- Upload the two or three documents customers ask about most — usually a price list and a policy document.
- Add five manual entries for the questions your staff answers daily from memory.
- Test like a difficult customer. Ask about prices, hours, refunds, delivery edge cases. Note anything wrong.
- Fix at the source. Correct the knowledge entry, not the individual chat.
- Turn on handover rules so anything uncertain reaches a human.
Keeping it fresh
Your business changes — prices, seasonal menus, holiday hours. Treat the knowledge base like your storefront sign: update it the moment reality changes, because the bot repeats it hundreds of times a week. Assign one owner; ambiguity about who updates the bot is how bots rot.
Frequently asked questions
Do I need technical skills? No — connecting a website and uploading documents is the whole job.
How much content do I need? Less than you think. Twenty well-chosen answers cover the vast majority of real questions for most small businesses.
Can the bot answer in other languages? Yes — with multilingual mode, the bot detects the customer's language and answers in it, using the same knowledge.
What about questions I don't want automated? Configure handover keywords and categories — refunds, complaints and custom quotes typically go straight to humans.
Train it once, and it answers on every channel you connect. Start free and see what your bot can already answer after ten minutes.