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

Writing Content AI Can Retrieve: A Practical Guide

AI SNS CHAT EditorialPublished 7 min read

Learn how to structure help articles, FAQs and policies so an AI agent finds the right passage and answers customers accurately on chat channels.

Most wrong answers from a support AI are not a model problem. They are a writing problem: the answer exists in your documents, but it is buried, vague, or split across three pages that never mention each other. This guide shows how to write and structure knowledge base content so an AI agent can find the right passage and reply correctly on website chat, WhatsApp and Messenger.

How an AI agent actually finds your content

When a customer asks a question, the agent does not read your whole knowledge base. It searches for the passages that look most relevant to the question, then writes a reply based on what it found. Your content is cut into pieces, and those pieces compete to be selected.

That has three practical consequences:

  • A passage must make sense on its own. The agent may see one paragraph without the page title or the paragraph before it.
  • Wording matters. If customers say "money back" and your page says only "remittance reversal," the match is weaker.
  • Only what is written can be retrieved. If the answer lives in a staff member's head, or inside an image, the agent cannot use it.

Think of each paragraph as a card in a deck. The agent draws a few cards, and each one has to carry its own meaning.

Write one topic per section, and name it like a customer would

The single most useful habit is to give each section exactly one job and a heading that mirrors the question a customer types.

Use question-style headings

Compare these two headings:

WeakStrong
ReturnsCan I return an item I bought on sale?
Shipping infoHow long does delivery take to Singapore?
AccountHow do I change the email on my account?

The strong versions match how people actually write in a chat window, which is short, informal and specific.

Keep each section self-contained

Avoid "as mentioned above," "see the previous section," or "the same rule applies." If the agent retrieves only that paragraph, the reference points to nothing. Repeat the key fact instead.

Put the answer first, then the conditions

Chat replies are short. Your source text should lead with the direct answer, followed by exceptions and detail.

Harder to retrieve and use:

"Our returns process has evolved over time and depends on several factors, including product category, purchase date and the condition of the goods. In general, customers may be eligible for a refund."

Easier:

"Yes, you can return unused items within 30 days of delivery for a full refund. Sale items can be exchanged but not refunded. To start a return, reply with your order number."

The second version gives the agent a clear answer, the boundary conditions, and the next step. (The 30-day figure here is an example for illustration, not a recommendation.)

A simple answer pattern

  1. State the answer in one sentence.
  2. Give the conditions or exceptions.
  3. Say what the customer should do next.
  4. Name who to contact if it does not apply.

Use the customer's words, and add the synonyms

Customers rarely use your internal vocabulary. Someone on WhatsApp may write "cancel my sub," "stop billing," or "unsubscribe." If your document says only "terminate your subscription," retrieval has less to grab.

Do this deliberately:

  • Use the plain term as the main wording and mention the formal term once.
  • List common alternate phrasings in a short line such as "Also called: cancel, stop, unsubscribe."
  • Spell out abbreviations once: "cash on delivery (COD)."
  • Keep product and plan names exactly consistent. Do not call the same plan "Pro," "Professional" and "the top tier" on different pages.

If you serve several languages, write the source content cleanly in each language you intend to support rather than relying on a single dense English page to cover everything. Mixed-language paragraphs are harder to match.

Make facts explicit: dates, units, names and limits

Vague text produces vague answers, and invented ones when the agent tries to fill the gap. Replace soft language with checkable facts.

VagueExplicit
Delivery is quickStandard delivery takes 3 to 5 business days within the US
A small fee appliesA $5 handling fee applies to orders under $50
Contact us during office hoursSupport is available Monday to Friday, 9:00 to 18:00 Japan Standard Time
Newer plans include thisAvailable on the Standard plan and above

(All values are placeholders for illustration. Use your real ones.)

Also state what you do not do. "We do not ship to PO boxes" is a retrievable fact. Silence is not. An agent without a clear negative may guess.

Hard for AI to retrieve
Easy for AI to retrieve
"Refunds are handled per our standard policy, see above."
"Refunds are issued to the original payment method within 7 business days of approval."
"Contact support for anything unusual."
"For damaged items, send a photo in this chat and we will arrange a replacement."
"Hours may vary."
"Chat support: Monday to Friday, 9:00 to 18:00 JST."
Same policy, written two waysIllustrative example

Format for scanning: short blocks, tables and lists

Structure helps both people and retrieval.

  • Short paragraphs of two to four sentences keep each chunk focused.
  • Lists work well for steps and requirements. Number steps that must happen in order.
  • Tables suit comparisons such as plans, regions or fees. Add a sentence above the table saying what it covers, because a table with no context is hard to interpret.
  • Avoid text locked in images. A pricing screenshot or a PDF scan of a poster may not be readable. Write the same content as text.
  • Remove duplicates and outdated pages. Two pages that disagree about the return window will produce inconsistent replies. Keep one source of truth per topic.

Handle exceptions in their own place

Do not bury an exception in the middle of a long paragraph. Give it a labelled line: "Exception: items marked Final Sale cannot be returned." Short, labelled exceptions are far easier to surface at the right moment.

Test with real questions and fix what fails

Writing for retrieval is iterative. The most reliable method is to test with questions that customers really ask.

  1. Collect 20 to 30 real questions from past chats, emails or reviews, in the customers' own wording.
  2. Ask each one and read the reply critically: right answer, wrong answer, or "I don't know"?
  3. For each miss, find out why. Was the content missing, unclear, in the wrong section, or contradicted by another page?
  4. Fix the source text, not just the single reply, then re-test.
  5. Repeat after any policy change.

Track patterns. If many misses involve the same topic, the content on that topic probably needs restructuring rather than another patch.

  1. 01Collect real questions
  2. 02Test the agent's answers
  3. 03Fix the source text
  4. 04Re-test after changes
A review loop for knowledge base contentIllustrative example

Doing this in AI SNS CHAT

AI SNS CHAT gives you several places to put knowledge, and each rewards the habits above.

Add your sources. In the sidebar, open Data sources. You can add Files, Text snippets, a Website, Q&A pairs, Notion pages, Social content, Tickets and Suggestions. If your documentation already lives elsewhere, the connectors under Channels can pull from Confluence, GitBook, ReadMe, Guru, Zendesk Guide, Google Drive, OneDrive, Dropbox, Box, SharePoint and others. Wiki-type connectors read up to 1.5 million characters and 400 pages per sync, and the product tells you if content was truncated, so check that your most important pages were included.

Use Q&A and text snippets for precision. For policies that must be exact, such as return windows or opening hours, a short Q&A entry or text snippet written in the answer-first pattern is easy to control and easy to update.

Test in Playground. Use Playground to ask your collected real questions before customers do. Playground also offers an A/B model comparison, useful when you want to see whether a wording problem or a model choice explains a weak answer. Model settings are under Agent → Settings → AI.

Teach the agent from real chats. Conversations on most channels appear in Activity → Chat logs, while email sits in Inbox. In both, you can take over a chat and edit a reply into the ideal answer to train the agent. Thumbs up and down ratings in Chat logs help you find the answers that need work. Analytics, including Topics and Question history, shows what customers keep asking about, which tells you where to write next.

Let the agent ask for what it needs. In Contacts and Custom attributes, the description of each field tells the agent when to ask for it. Write those descriptions as plainly as your articles.

Keep control of edits. Workspace settings → Activity records each create, edit, delete or upload, and each data source shows who added it and who last edited it. That makes it easier to trace when a changed document altered an answer.

Note that knowledge-base passages shown to strangers are scrubbed for personal information, but you should still avoid putting private customer data in public-facing articles. For more on capabilities, see features, and for plan limits such as storage, see pricing.

Conclusion

Retrieval-friendly writing is mostly good support writing: one topic per section, answer first, the customer's own words, explicit facts and no contradictions. Test with real questions, fix the source, and repeat. Over time, your knowledge base becomes something both customers and AI can rely on.

Frequently asked questions

How long should each knowledge base article or section be?

Aim for sections that cover a single question in roughly two to six short paragraphs or a compact list. If a section answers more than one question, split it. There is no magic number, but a passage should always be understandable without the text around it.

Should I write the same content in several languages?

If you serve customers in more than one language, it is better to maintain clean source text in each language you support than to mix languages in one page. The welcome message in AI SNS CHAT is auto-translated for the languages the widget speaks, but your policy content is only as clear as what you write.

What if the agent still gives a wrong answer after I rewrite the content?

Look for a conflicting page, an outdated source, or a missing fact first. Test the same question in Playground, fix the source, and ask again. If the content is clear and the answer is still weak, compare models with the A/B tool.

Can the agent answer from the live web if my knowledge base lacks an answer?

Web answers are limited. The built-in fallback applies only to LINE WORKS, and the Tavily action can give any agent real-time web search. For most channels the safest approach is to write the missing answer into your knowledge base.

#knowledge base#content writing#AI retrieval#customer support

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.