Skip to content
AI-SNS.IO — CHAT SUPPORTAI-SNS.IO — CHAT SUPPORT
Knowledge Base

Building a Knowledge Base AI Can Use to Answer Correctly

AI SNS CHAT EditorialPublished 6 min read

How to structure, write and maintain a knowledge base so your support AI gives accurate answers on website chat, WhatsApp and Messenger.

Most wrong answers from a support AI are not the model's fault. They come from a knowledge base written for humans who skim, guess and ask a colleague when something is unclear. An AI does none of that: it reads what you gave it and answers from it.

This guide shows how to build a knowledge base that an AI can use reliably: what to include, how to write it, how to test it and how to keep it current.

Start with the questions, not the documents

The usual instinct is to upload every PDF, policy and help article you own. That produces a large pile, not a useful base. Start from what customers actually ask.

  1. Export or read a few weeks of real conversations from website chat, WhatsApp and Messenger.
  2. Group them by topic: shipping, returns, pricing, account access, setup, opening hours.
  3. Rank topics by how often they appear.
  4. For each of the top topics, write down the exact answer a good agent would give today.

If a topic has no clear answer in your company, the AI will not invent a better one. Settle the answer first, then write it down.

Decide what the AI should not answer

Equally important is the boundary. List the topics that must go to a person: refunds above a certain value, legal complaints, medical or safety questions, account-specific disputes. A knowledge base that is silent on these, combined with a clear handoff rule, is safer than one that offers half-answers.

Write for retrieval: one idea per passage

AI systems typically find relevant passages and answer from them. That means each passage should make sense on its own, without the reader having seen the rest of the page.

Practical writing rules

  • One topic per section. A section titled "Returns" should not also explain delivery delays.
  • Put the question in the heading. "How long does standard delivery take?" matches how customers phrase things.
  • State the answer first, then conditions and exceptions.
  • Name things in full. Write "the Standard plan" rather than "it" or "this tier", because the passage may be read in isolation.
  • Include the limits. Dates, regions, minimum amounts and eligibility belong in the same passage as the rule.
  • Avoid pointing elsewhere. "See the section above" means nothing to a retrieved fragment.

Before and after

Written for skimming humans
Written for AI retrieval
"Returns: see policy. Usually 30 days, some exceptions."
"Can I return an item? Yes, within 30 days of delivery if unused and in original packaging. Sale items and gift cards cannot be returned."
"Shipping is quick in most areas."
"How long does delivery take? Domestic orders arrive in 3 to 5 working days. International orders take 7 to 14 working days."
"Contact us for help."
"How do I reach a person? Ask to speak to a person in this chat, and a team member will take over during staffed hours."
Writing the same policy for people versus for AIIllustrative example

The right-hand versions are longer, but every word is doing work. The AI no longer has to guess what "usually" means.

Handle conflicts, gaps and stale content

Contradictions are the most common cause of confident but wrong replies. If one page says returns take 14 days and an older one says 30, the AI may quote either.

A short audit routine

  • Search your sources for the same fact (price, deadline, address) and make sure it matches everywhere.
  • Delete or archive old versions instead of leaving them "just in case".
  • Date-stamp time-sensitive content such as promotions, and remove it when it ends.
  • Mark each document with an owner who is responsible for it.

Be explicit about "I don't know"

Add a line to the agent's instructions about what to do when the knowledge base has no answer: say so plainly and offer a human handoff. An AI that admits a gap is far more trustworthy than one that fills it.

Choose the right source for each kind of content

Not all knowledge is the same shape. Matching the format to the content makes upkeep easier.

Content typeGood formatWhy
Short, exact answers (hours, fees, policies)Question and answer pairsPrecise, easy to review
Long guides and manualsFiles or a connected documentation toolStays in sync with the original
Public product informationWebsite crawlReuses what is already published
Frequently changing internal notesWiki or document connectorOne place to edit
Small clarifications and rulesText snippetsQuick to add and correct

A practical pattern is to use your documentation for depth and a small set of hand-written Q&A pairs for the top questions, where wording must be exact.

Test with real questions before you go live

A knowledge base is only proven when someone asks it questions. Build a test set and run it every time you make a significant change.

Building the test set

  • Take 30 to 50 real customer questions, including badly spelled and informal ones.
  • Add questions in other languages if you serve them.
  • Include a few the base should not answer, to confirm the AI declines or hands off.
  • Record the expected answer for each.
  1. 01Ask test questions
  2. 02Compare with expected answers
  3. 03Fix the source passage
  4. 04Retest and go live
A simple improvement loopIllustrative example

When an answer is wrong, resist the urge to patch the AI's instructions first. Ask which passage it used, and whether that passage is unclear, missing or contradicted elsewhere. Fixing the source fixes every future variation of the question.

Keep it alive after launch

Launch is the start of the work. Real conversations reveal gaps no planning session finds.

  • Review a sample weekly. Read a handful of conversations, especially ones rated badly or handed to a person.
  • Turn corrections into content. When a human improves an answer, make sure that improvement lands in the knowledge base.
  • Track change. Know who edited what and when, so errors can be traced and reversed.
  • Update on every business change. New prices, new products and new policies should have a checklist item for "update the knowledge base".

For example, if a shop receives 300 chats a month and reviews 20 of them each week, it sees roughly a quarter of its conversations every month. That is enough to spot recurring gaps without turning review into a full-time job.

Doing this in AI SNS CHAT

AI SNS CHAT gathers knowledge sources in one place. Open Data sources in the sidebar and you will find Files, Text snippets, Website, Q&A, Notion, Social, Tickets and Suggestions.

Choosing sources. Use Q&A for the exact answers you wrote in the first step, Files for manuals, and Website to read pages you already publish. Under Channels, connectors read from tools such as Confluence, GitBook, ReadMe, Guru, Zendesk Guide, Google Drive, OneDrive, Dropbox, Box, SharePoint and Notion, so your documentation can stay where your team already edits it. Wiki-style connectors read up to 1.5 million characters and 400 pages per sync, and AI SNS CHAT reports when content was truncated, so you know if something was left out. Storage and source limits depend on your plan, and some limits are set by the platform administrator.

Testing. Use the Playground to ask your test questions before connecting a channel. Under Agent, Settings, AI you can choose the model; advanced models and model fallback need the Hobby plan or higher.

Improving from real chats. Conversations from website chat, WhatsApp, Messenger and other channels appear under Activity, Chat logs, where you can filter, flag, and review thumbs up and down ratings. Email conversations are in Inbox. In both places, a person can take over a chat, which pauses the bot, and can edit a reply into the ideal answer to train the agent, which turns a correction into lasting knowledge.

Handoff. Under Helpdesk, Scheduling (Standard or higher) you set shifts per person and timezone, and handoffs go to whoever is on shift. AI SNS CHAT adds the "what happens next" line for the customer from the rota and never invents a wait time.

Accountability. In Workspace settings, Activity, owners and admins can see every create, edit, delete or upload with name, email, feature and time. Each data source also shows who added it and who last edited it. If you want to try changes safely, agents have restore points for rollback.

Privacy note: knowledge-base passages shown to strangers are scrubbed for personal information. Details are on the security page, and plan details are on the pricing page. You can see the wider list of capabilities under features.

Frequently asked questions

How much content does an AI need to answer well?

Less than most teams expect, if it is the right content. Cover your most frequent questions with clear, specific answers first, then expand based on gaps you see in real conversations.

Should I upload everything I have?

No. Old, duplicated or contradictory documents cause wrong answers. Upload current material, remove superseded versions, and give each source an owner.

What should the AI do when the knowledge base has no answer?

It should say it does not have that information and offer to connect the customer with a person. Write this into the agent's instructions and set up a handoff so there is someone to receive the chat.

How often should I review the knowledge base?

Review a sample of conversations weekly at first, and update content whenever prices, policies or products change. After a few months, a lighter monthly review is often enough, depending on how fast your business changes.

#knowledge base#AI support#content quality#chatbot training

Put this into practice with AI SNS CHAT

Build an AI agent trained on your own content and connect it to your website, LINE, WhatsApp and more. Free to start, no credit card required.