Chapter 1 · The scenario
Example: The return window that changed twice
A Jaipur saree retailer extends its return window from 7 to 10 days for the festive season. The website is updated, but the bot's sources still include an old policy PDF, so the AI agent tells some customers 7 days and others 10. The fix is structural: one "Return window" article owned by the operations lead, with an effective date and a review date of 15 January, the PDF removed from the bot's sources, and the support team's saved reply rewritten to match. A weekly review of unanswered questions then adds two new articles the bot kept failing on — cash-on-delivery limits by pincode and blouse-stitching timelines.
Chapter 2 · What is a knowledge base?
A knowledge base (KB) is the curated body of content a business answers customers from: FAQs, help articles, return and refund policies, delivery coverage, price lists, product specifications and internal procedures written for customer-facing use. On WhatsApp the same knowledge base can feed a chatbot, an LLM-powered AI agent through retrieval-augmented generation (RAG), the suggestions an agent assist tool shows a human, and the help pages customers read on the web.
It is the content, not the machinery around it. A vector database stores embeddings of knowledge-base chunks so an AI system can find the relevant passage quickly; it does not decide whether the passage is correct. Canned responses are short reply snippets an agent inserts with a shortcut; they are outputs, while a knowledge base is the source of truth those snippets — and every AI answer — should agree with.
Because AI systems quote it with confidence, a knowledge base needs more discipline than a traditional help centre. Each article should answer one question, carry an owner, an effective date and a next-review date, and make sense when read on its own, since retrieval hands the model a chunk rather than the whole page. Two articles that disagree, or one that quietly went out of date, are the most common source of the wrong answers people blame on hallucination.
Key takeaways
- A knowledge base is the curated content — FAQs, policies, product facts — that bots, AI agents and human agents answer from.
- An AI answer can only be as current and consistent as the knowledge base behind it; stale articles become confident wrong answers.
- Write one question or policy per article, with an owner, an effective date and a review date.
- The knowledge base is the content; the vector database only indexes it, and canned responses are short reply snippets.
- Log every question the bot could not answer and turn the recurring ones into new articles each week.
Chapter 3 · How it unfolds
How does a knowledge base work?
Inventory the sources of truth
List where answers live today — the website, policy PDFs, the support team's saved replies, the product catalogue — and decide which one wins for each topic. Retire the copies.
Write atomic articles
One question or policy per article, phrased the way customers ask it, answer first, then conditions and exceptions. Add metadata: owner, effective date, review date, audience, product, region and language.
Shape it for retrieval
Use clear headings, keep each section self-contained (repeat the product or policy name instead of "see above"), and convert tables and price lists in images or scanned PDFs into plain text.
Connect it to every channel
Index it for the bot or AI agent (chunks and embeddings for RAG), expose it to agent assist in the inbox, and publish the public parts as help pages — all from the same source.
Own it and keep it fresh
Review on a schedule and on every trigger — a price change, a new courier, a policy update. Feed the log of unanswered or escalated questions back in as new articles.
Chapter 4 · In code
A knowledge base in code
{"id": "kb-returns-window","question": "How many days do I have to return a saree?","answer": "You can return unworn sarees within 10 days of delivery. Stitched blouses and sale items marked Final Sale cannot be returned.","details": "Start a return by replying RETURN on WhatsApp. Pickup is free in serviceable pincodes; refunds reach the original payment method within 5–7 working days of pickup.","owner": "operations-lead","effective_from": "2026-10-01","review_by": "2027-01-15","audience": "customer","languages": ["en", "hi"],"channels": ["bot", "agent_assist", "help_centre"],"status": "published","supersedes": "kb-returns-window-v3"}
Metadata turns a paragraph into something you can own, expire, filter by channel and audit when a customer disputes an answer.
Chapter 5 · Why it matters
Why does a knowledge base matter?
Every automated answer on WhatsApp is only as good as the text behind it. A bot that quotes last year's return window or a delivery promise for a pincode you no longer serve creates refunds, escalations and complaints at machine speed. A maintained knowledge base is what lets a business let AI reply without fear, gives human agents the same answers the bot gives, and turns "what did we tell this customer?" into a question with one answer.
Chapter 6 · Where it goes wrong
Common mistakes with a knowledge base
Uploading everything and hoping
Dumping every PDF, old newsletter and scanned brochure gives the AI contradictions to choose from. Curate: include only current, customer-safe sources and retire superseded versions.
No owner, no review date
Articles without an owner drift out of date silently. Assign each one to a person or team and alert them when the review date passes or a related policy changes.
Writing for readers who scroll
"As mentioned above" means nothing in a retrieved chunk. Make each section stand alone and name the product, plan or policy it applies to.
Mixing internal notes into bot sources
Discount authority, escalation contacts or fraud rules in a bot-facing article can be repeated to customers. Keep internal guidance in a separate, agent-only collection.
Chapter 7 · What good teams do
Best practices for a knowledge base
- Phrase article titles as customer questions, including the Hinglish or regional wording they actually use.
- Keep a changelog per article so you can show what the bot was told on any given date.
- Measure answer coverage: the share of incoming questions with a matching article.
- Translate the articles customers need most rather than relying on on-the-fly translation for policy text.
- Review the top escalated topics every week — they are your next articles.
Chapter 8 · The facts
A knowledge base: key facts and limits
| Unit of content | One question or policy per article |
|---|---|
| Minimum metadata | Owner, effective date, review date, audience, language |
| Review triggers | Price, policy, courier or product changes — plus a fixed schedule |
| Meta Business Agent sources | Business information, FAQs, websites and files |
| Retrieval unit for RAG | Chunks — typically a section, not a whole document |
Chapter 9 · Knowledge Base FAQs