To “train” a chatbot on your website and PDFs, you don’t retrain an AI model. You give a retrieval-based chatbot your pages and documents; it splits them into passages, finds the passages relevant to each question, and writes an answer from them. The quality of the answers depends mostly on the quality of that content: clear, current, text-based documents with one topic per section produce good answers, while scanned PDFs, stale pages and contradictory documents produce bad ones.
This guide covers how the process works, how to prepare your content, what breaks retrieval and how to test the result.
How “training” on your content actually works
Most modern website and document chatbots use an approach called retrieval-augmented generation (RAG). In plain English:
- Ingest. The system reads your website pages and uploaded documents and extracts the text.
- Chunk. It splits that text into smaller passages, often a few paragraphs each.
- Index. Each passage is converted into a numeric representation of its meaning (an “embedding”) and stored so it can be searched by meaning, not just keywords.
- Retrieve. When a visitor asks a question, the system finds the passages most relevant to it.
- Answer. A language model writes a response using those passages, ideally citing which ones it used.
The key point: the bot can only be as accurate as the passages it retrieves. If the right passage doesn’t exist, is hard to find, or conflicts with another passage, the answer suffers. That’s why content preparation matters more than any setting.
Step 1: Inventory your knowledge sources
Before uploading anything, list what the bot should know and where it lives.
| Source | Examples | Notes |
|---|---|---|
| Website pages | Services, pricing pages, about, contact, blog posts | Check which pages are current |
| PDFs | Brochures, policies, manuals, price lists | Check whether they’re text or scanned images |
| Word documents (DOCX) | Internal FAQs, procedures, onboarding guides | Often the most up-to-date source |
| Markdown | Product docs, help center exports, knowledge bases | Usually clean and well-structured |
| Unwritten knowledge | Answers that only exist in staff heads | Needs to be written down first |
Then pull a list of the 30 to 50 questions visitors actually ask, from support emails, contact forms, call notes and sales conversations. Mark which source answers each one. Every question with no source is a content gap.
Step 2: Prepare your website content
- Remove or exclude outdated pages. Old promotions, discontinued services and archived news posts will be retrieved and quoted as if they’re current.
- Put key facts in text, not images. Opening hours in a banner image or prices in an infographic are invisible to text extraction.
- Make each page about one thing. A page that covers returns, shipping and warranty in one block of text is harder to retrieve from than three clear sections.
- Use descriptive headings. “Refund policy for annual plans” retrieves better than “More info.”
- Watch out for content behind logins, tabs or scripts. Content that only appears after a click or is loaded by JavaScript may not be read by every crawler. Test whether it’s picked up.
Step 3: Prepare your documents (PDF, DOCX, Markdown)
Structure documents for retrieval
- Use real headings (Heading 1, Heading 2 in Word;
##in Markdown), not bold text pretending to be a heading. - Keep one topic per section, and restate the subject in each section. “The annual plan can be cancelled within 30 days” is better than “It can be cancelled within 30 days,” because a retrieved chunk may be read without the paragraph before it.
- Break very long documents into logical files, for example one per product or policy.
- Add a short summary at the top of long documents.
Write a dedicated FAQ document
The single most effective improvement is usually a plain FAQ document written for the bot: each question as a heading, followed by a direct two-to-four-sentence answer. It fills gaps and gives the bot clean, retrievable passages for the most common questions.
Name files clearly
If your chatbot shows citations, visitors will see the document names. “Returns-Policy-2026.pdf” is more trustworthy than “final_v3_USE_THIS.pdf.”
What breaks retrieval
These are the most common causes of wrong or missing answers:
| Problem | Why it breaks | Fix |
|---|---|---|
| Scanned PDFs | The PDF is an image; there’s no text to extract | Run OCR, or re-export from the original file |
| Complex tables | Rows and columns can lose their structure when extracted as text | Convert key tables into sentences or simple lists |
| Stale pages | Old content is retrieved and presented as fact | Remove, update or exclude it |
| Contradictory documents | Two sources give different answers; the bot may pick either | Keep one source of truth per topic |
| Vague headings and pronouns | Chunks lose context when separated from surrounding text | Use specific headings; restate subjects |
| Content only in images or video | No text to index | Add text versions or transcripts |
| Huge catch-all documents | Relevant passages compete with lots of noise | Split by topic |
| Headers, footers and boilerplate | Repeated text clutters every chunk | Remove from documents where possible |
Contradictions deserve special attention. If your website says “free returns within 30 days” and a 2023 PDF says “14 days,” the bot has no way to know which is correct. Find and resolve these before launch.
Step 4: Test the answers
Testing isn’t optional. Build a question set that covers:
- Your top questions, asked the way customers phrase them, including typos and casual wording
- Questions answered only in documents, to confirm PDFs and DOCX files are being used
- Questions with no answer in your content, to confirm the bot doesn’t invent one
- Out-of-scope questions, such as general trivia or competitor questions
- Questions where sources might conflict
For each answer, check whether it’s correct, whether the cited source is the right one, and whether the tone fits your brand. When an answer is wrong, look at the source first: more often than not, the fix is content, not configuration.
For a full test matrix, see our pre-launch chatbot testing checklist, and for how to judge answers using citations, read about AI chatbots with source citations.
Step 5: Keep the knowledge current
A chatbot trained on last year’s content will confidently give last year’s answers. Build a simple routine:
- When a page or policy changes, update or re-upload the source.
- Review conversation history regularly for questions the bot handled poorly.
- Write new FAQ entries for recurring questions that have no good source.
- Remove documents that are superseded, rather than adding a newer version alongside them.
How this works in Techvia AI Bot
In Techvia AI Bot, a bot’s knowledge comes from website content plus PDF, DOCX and Markdown documents. When a visitor asks a question, the bot retrieves the relevant content and answers with source citations, so you can see exactly which page or document an answer came from. Conversation history lets you review real questions after launch and spot gaps to fill. Agencies can keep each client’s knowledge sources in a separate workspace.
The preparation advice above applies whatever platform you use: clean, current, text-based content is what makes any retrieval chatbot accurate.
Frequently asked questions
Do I need to retrain an AI model to make a chatbot use my documents?
No. Retrieval-based chatbots search your content at the moment a question is asked and answer from what they find. Updating the content updates the answers; no model training is involved.
Can a chatbot read scanned PDFs?
Only if the text can be extracted. A scanned PDF is an image of a page, so it needs OCR (optical character recognition) first, or you should use the original digital file instead. Check with your platform how it handles scanned documents.
How many documents can I add?
That varies by platform and plan. More important than the count is quality: fewer, well-structured, non-contradictory documents usually outperform a large pile of overlapping ones.
Why does my chatbot give an outdated answer?
Usually because an outdated page or document is still in its knowledge sources. Check the citation on the answer, then remove or update that source.
Next step
If you’d like to see a chatbot answer from your own website and documents with citations, book a demo of Techvia AI Bot and bring a few real questions to try.