How to Build a FAQ Knowledge Base for Your AI Chatbot
The knowledge base is the real product in any AI support project. Here is how to organize yours so the bot answers accurately and escalates correctly.
Your chatbot is only as good as its knowledge base. Models do not magically know your return window, warranty language, or which SKUs are final sale. This article is a practical structure for FAQs, policies, and product info that produce accurate LLM answers.
Start from real demand, not a blank template
Export 30–90 days of support tickets, chat logs, and Instagram DMs. Cluster them into themes. The top clusters become your first knowledge articles. Ignore rare one-off edge cases for v1.
Article structure that models handle well
- Clear title that matches how customers ask
- One short direct answer in the first sentence
- Details, exceptions, and links below
- Last-reviewed date and owner
Avoid dumping a 20-page PDF and hoping retrieval finds the right paragraph. Chunk content by intent.
Core knowledge categories
- Orders, shipping, tracking
- Returns, exchanges, refunds
- Products, sizing, materials
- Accounts, payments, promos
- Hours, locations, booking
- Escalation and “we cannot help with” boundaries
Grounding rules
Decide what the bot may never invent: prices if dynamic, legal advice, medical advice, custom discounts, and competitor comparisons. Write refusal language that still feels helpful — for example, collect order email and escalate.
Maintenance cadence
Assign one owner. Review failed answers weekly for the first month, then biweekly. When a policy changes, update the source article the same day you update the website.
We use this process on LLM customer service pilots. If you need a readiness check first, read Is your business ready for an AI chatbot?
Example article: “How long does shipping take?”
Direct answer: Standard shipping arrives in 5–7 business days after your order ships. Most orders ship within 1–2 business days.
Details: Cutoff times, weekend processing, holiday delays, and international exclusions. Link to the shipping policy page. Escalation: if tracking has not moved for 5 business days, collect order email and route to a human.
Write every high-volume topic this way. Short answers retrieve cleanly; long digressions make the model more likely to miss the point.
Tools and formats
A shared Notion or Google Doc works for v1. Helpdesk articles work better once you have volume. Whatever you choose, one source of truth beats five outdated PDFs. When the website policy and the knowledge base disagree, the bot will eventually embarrass you — keep them synced.