Learn how to prepare PDFs and manuals so your support AI answers accurately on web chat, WhatsApp and Messenger, and how to test and maintain it.
You have a 200-page product manual, a stack of PDF spec sheets and a returns policy last updated three years ago. You upload them all to an AI agent and the answers come back vague, outdated or just wrong. The problem is rarely the AI. It is usually the documents, and how they were prepared, scoped and tested. This guide walks through what to fix before you upload, how to check the result, and how to keep it accurate over time.
Why documents that read well to humans trip up AI
A manual is written for someone who reads it in order, with a table of contents, diagrams and context built up over chapters. A support AI works differently. It retrieves a few relevant passages in response to a customer's question and writes an answer from those passages alone.
That creates a few common failure modes:
- Missing context. A passage that says "set it to the second option" means nothing without the heading three pages earlier.
- Contradictions. Two versions of a warranty policy sit in the knowledge base, and the AI picks either one.
- Scanned pages. If a PDF is an image of text, there may be no readable text to learn from at all.
- Tables and diagrams. Information locked in a picture or a complex table layout often gets garbled or ignored.
- Marketing fluff. Brochures are full of claims and short on the details customers actually ask about.
Understanding this changes how you prepare content. You are not handing over a book. You are building a set of passages that each make sense on their own.
Step 1: Decide what the AI should know, and what it should not
More documents do not mean better answers. Before uploading anything, sort your material into three groups.
| Group | Examples | Action |
|---|---|---|
| Customers ask about it often | Setup guides, pricing rules, shipping, returns, troubleshooting | Upload, and clean carefully |
| Rarely asked, but risky if wrong | Warranty terms, safety notes, compliance statements | Upload, and review wording line by line |
| Internal or outdated | Old price lists, staff-only procedures, draft documents | Leave out |
A good starting test: look at the last 50 questions your team answered on web chat, WhatsApp or Messenger. Which documents would have answered them? Start with those, not with everything you own.
Step 2: Check that the text is actually readable
Open each PDF and try to select a sentence with your cursor. If you cannot highlight individual words, the file is a scan.
- Scanned PDFs: run them through OCR (optical character recognition) first, then proofread the output. OCR often confuses similar characters, such as 0 and O, or 1 and l, which matters in part numbers and model names.
- Multi-column layouts: text can be read across columns instead of down them, mixing two sentences into one. Spot-check a few pages.
- Headers and footers: repeated page numbers and legal lines add noise. Remove them if you can export the document from its original source.
- Images with text: a screenshot showing an error code is invisible as text. Add the error code and its fix as a sentence beneath it.
If you still have the original Word, Google Docs or Markdown file, use that instead of the exported PDF. Source files almost always keep cleaner structure.
Step 3: Restructure for retrieval
This is where most of the accuracy gain comes from. Reshape the content so each section can stand alone.
Write headings that name the topic
"Section 4.2" tells the AI nothing. "How to reset the router to factory settings" is exactly the phrase a customer might type.
Repeat the subject in each section
Instead of "It should be held for ten seconds," write "Hold the reset button on the X200 router for ten seconds." The second version still makes sense when it is retrieved by itself.
Turn dense tables into sentences
A specification table with twelve columns is hard to read reliably. Rewrite the important rows as short statements: "The X200 supports up to 64 connected devices and Wi-Fi 6."
Keep one answer in one place
If shipping times appear in four documents, consolidate them into one current source. Then retire the others.
Add a question-and-answer layer for the top questions
For your twenty most common questions, write a plain question and a short, approved answer. This gives the AI a precise, trusted source for the topics that matter most, and it sits alongside your longer manuals.
- 01Audit documents
- 02Clean and restructure
- 03Upload and scope
- 04Test with real questions
Step 4: Make the content say what you want it to say
Tone and boundaries come from the agent's instructions, not from your manual. Use the documents for facts and the persona for behaviour.
- Tell the agent to answer only from your documents for product and policy questions.
- Tell it what to do when the answer is not there: for example, offer to pass the chat to a person rather than guess.
- Keep warranty, refund and safety language close to your source text, so the AI does not paraphrase a legal term into something new.
An honest "I don't have that information, let me connect you with the team" is far better than a confident wrong answer about a refund.
Step 5: Test with real questions before customers do
Never judge a knowledge base by asking the three questions you already know the answer to. Build a test list instead.
- Collect 30 to 50 real questions from past chats and emails.
- Include awkward ones: typos, vague phrasing, two questions in one message.
- Include questions your documents do not answer, to see whether the AI admits it.
- Run them all and mark each answer correct, partly correct, wrong or invented.
- For every miss, find out why: missing content, poor structure, or conflicting sources.
Fix the source, not just the symptom. If the AI got a return window wrong, the answer is usually to correct or remove a conflicting document rather than to add yet another one.
Step 6: Keep it current
Manuals change. Products are revised, prices move and policies are updated. A knowledge base that was perfect in March can be misleading by September.
- Give each source an owner who is responsible for it.
- Review high-risk content (warranty, pricing, safety) on a fixed schedule.
- When a product is discontinued, remove its documents, rather than letting them linger.
- Review chats where customers gave a thumbs-down or where a person had to step in. These point straight to gaps.
- When a human agent writes a better answer, feed it back so the AI learns from it.
For example, if your team reviews 20 flagged chats each week and finds that most trace back to two or three outdated pages, you know exactly where to spend your time.
Doing this in AI SNS CHAT
AI SNS CHAT gives you several ways to bring documents in, so you can match the method to the material.
Adding your content. In the sidebar, open Data sources. You will find Files (for PDFs and manuals), Text snippets, Website, Q&A, Notion, Social, Tickets and Suggestions. Use Files for your cleaned manuals, Q&A for your approved answers to the top questions, and Text snippets for short facts such as opening hours or a one-line policy.
Connecting living documentation. If your manuals already live in a documentation tool, you do not need to export PDFs at all. Under Channels, connectors are available for Confluence, GitBook, ReadMe, Guru, esa.io, DocBase, Qiita Team, Zendesk Guide, Google Drive, OneDrive, Dropbox, Box, SharePoint, Notion and an Obsidian vault via Git. Wiki-style connectors read up to 1.5 million characters and 400 pages per sync, and AI SNS CHAT tells you if anything was truncated, so a very large manual set may need to be split.
Mind your storage. Uploaded files count toward your plan's storage: 30 MB on Free, 200 MB on Hobby, 500 MB on Standard, 1 GB on Pro and 3 GB on Corporate. Cleaning and removing duplicates helps here as well as with accuracy. Extra storage can be added as a monthly add-on.
Setting behaviour. Use the agent's persona and guardrails to say "answer from the documents and hand off if unsure." Pair that with human handoff so a person can take over a chat when the AI reaches its limit.
Testing. Use the Playground to run your test list before connecting any channel. Once live on the chat widget, WhatsApp or Messenger, review conversations under Activity → Chat logs, where you can filter, flag and read thumbs up or down ratings.
Improving over time. When you take over a chat, you can edit your reply into the ideal answer to train the agent. The workspace activity log under Workspace settings shows who added or last edited each source, which helps with ownership.
Safety. Knowledge-base passages shown to strangers are scrubbed for personal data, but you should still keep sensitive files out. See /security for how practices are designed with reference to GDPR and APPI, and /pricing for plan limits.
Frequently asked questions
Should I upload PDFs or connect my documentation tool?
If your manuals already live in a tool such as Confluence, GitBook, Google Drive or SharePoint, connecting it is usually better because changes can be synced instead of re-uploaded. Use PDF upload for documents that exist only as files.
Why does the AI ignore information that is clearly in my PDF?
The most common causes are a scanned file with no readable text, information stored inside images or complex tables, or a heading structure that makes the passage meaningless on its own. Run OCR if needed, rewrite key tables as sentences and test again in the Playground.
How do I stop the AI from guessing when the answer is not in my manuals?
Write this into the agent's instructions: answer only from the provided documents and offer a handoff to a person when unsure. Then test with questions you know are not covered, to confirm it behaves that way.
How often should I update the knowledge base?
Review high-risk content such as pricing, warranty and safety on a regular schedule, and update any source as soon as the underlying product or policy changes. Chat logs with poor ratings are a useful signal of what to fix first.


