How to train a chatbot on your own PDFs, documents and website
Updated: · 8 min read
In short
To train a chatbot on your own PDFs and website, add the files and the site address to a builder, review what it understood, approve it, then test with real questions. Prepare first: use files with real text, one topic per file, current prices and dates, and no old contradicting versions. With Ravano the bot then answers only from what you approved and shows the source.
Most businesses already own the answers their customers keep asking for. They sit in a price list PDF, a policy page, a Word handbook or an old Telegram announcement. The work is not writing new content. It is getting those answers into a chatbot in a clean state, and checking them before a customer sees them. This guide covers the practical side: which sources Ravano accepts, how to prepare documents so the answers come out right, how the review step works, which test questions to ask, and what to do when the bot has no answer. It also explains in plain words how a bot can answer from your files without making things up.
What can you train a chatbot on?
Ravano accepts the places where your information already lives. You do not need to merge everything into one document. You can add several sources, and the bot treats them together once you approve what it learned.
| Source | Good for | Watch out for |
|---|---|---|
| Website address | Product pages, policies, FAQ, contact details | Old pages that still contradict newer ones |
| Price lists, catalogues, brochures, rules | Files that are only pictures of pages | |
| Word | Handbooks, procedures, long explanations | Tracked changes and several versions in one file |
| Excel or CSV | Products, plans, fees, schedules in rows | Unlabelled columns and empty cells |
| Telegram channel | Announcements, news, updated offers | Posts that are outdated by now |
| Product descriptions and announcements in captions | Informal wording that skips key details | |
| Free text | Facts written nowhere else | It is only as good as what you type |
How a bot answers from your files without guessing
A language model on its own answers from memory, and memory can be wrong. The approach used here is often called retrieval-augmented generation, or RAG. When a customer asks something, the system first looks up the most relevant passages in your approved information, then writes the answer from those passages and shows where they came from. The idea is the same as answering with the handbook open on the desk.
This is why the quality of the bot depends so heavily on your files. If the right passage exists, the answer can be accurate and carry a source. If it does not exist, the bot should say it does not know instead of filling the gap. No system removes every error, which is exactly why the review and test steps below matter. For a fuller explanation, read what is RAG.
Prepare your documents before you upload
Ten minutes of preparation usually improves answers more than any later tweak. Treat each file as something a new colleague will read and then be quizzed on.
Use files that contain real text
A PDF made by exporting from Word or a design tool holds text. A PDF made by scanning paper, or by taking photos of pages, is often just an image of text. Software that reads files may not be able to extract anything from an image-only page, so the bot would have nothing to learn from it. This is general advice that applies to most tools, not a statement about one feature.
A quick check: open the PDF and try to select a sentence with your mouse or search for a word. If you cannot, find the original Word or Excel file and upload that, or retype the important parts and add them as free text.
One topic per file
A file called shipping-and-returns will be understood better than a 90-page file called everything-final. Split large documents by topic: pricing, delivery, warranty, opening hours, how to book. Clear file names and clear headings inside each file help both you and the bot.
Show current prices and dates
Put the date or season on anything that changes: a price list, a promotion, opening hours over holidays. Write numbers in full with their unit and currency. A line such as "Delivery within the city: 2 to 3 working days" is useful. A line such as "fast delivery" is not.
Remove contradictory old versions
If last year's price list and this year's are both uploaded, the bot has two valid-looking answers. Delete or exclude the old one before you start. The same goes for a website that still has an old policy page live. Fix the source or leave it out; do not rely on the bot to guess which one is current.
Tips for your website, Telegram, Instagram and free text
- Website: add the address of the main site, then check that your policy, shipping and contact pages are up to date. A bot can only be as current as the pages it reads.
- Telegram channel and Instagram: these are full of offers that have expired. Treat them as supporting sources and look closely at the review step for old prices.
- Spreadsheets: give every column a clear heading, such as "Price (Toman)" or "Warranty (months)". Keep one row per item.
- Free text: use it for the small but important facts that appear nowhere else, such as "We do not deliver on public holidays". Write each fact as a short, complete sentence.
Review and approve: the step that makes it trustworthy
After you add sources, Ravano organises what it read and shows it to you in five sections. You go through each one, correct what is wrong and approve it. The bot answers only from the approved information, so this is where you decide what your customers will be told.
Do not skim. A few minutes per section is normal, and the order below catches most problems.
- Check every number first: prices, fees, deadlines, warranty periods, working hours.
- Look for statements that two sources disagree on, then keep the correct one and remove the other.
- Delete anything outdated rather than leaving it in as a harmless extra.
- Read the contact and hand-over information, because it is what customers will be sent to when the bot cannot help.
- Approve only when you would be comfortable seeing the text on a sign in your shop.
Test with real questions before you go live
Testing inside the panel is free, so use it generously. Write your questions down in advance, ideally from real customer messages, and check both the answer and the source under it.
Questions the bot should answer
Ask your ten most common questions, each in two or three different wordings. Customers rarely type the sentence you wrote in your FAQ. Confirm that numbers, conditions and exceptions come out exactly as approved.
Questions the bot should refuse
Ask something your files never mention, such as a product you do not sell. A good result is a clear statement that it does not know, not a plausible invention. Also try a question that depends on a detail you deliberately left out, to see whether the bot admits the gap.
Questions that expose conflicts
Ask about the one thing that changed recently, such as a price or a policy. If the answer mixes old and new, a stale source is still in the mix. Fix it and ask again.
What to do when the bot has no answer
Sooner or later a customer will ask something you did not cover. That is expected, and the right behaviour is an honest "I do not have that information", with a pointer to how to reach you. In Ravano the bot answers only from approved information and does not guess; the panel also lists questions it could not answer so you can see where your content has holes.
- If the answer exists in a file you did not add, add the file and approve the new content.
- If the answer exists only in your head, write it as a short free-text fact and approve it.
- If the answer is a case-by-case decision, such as a custom quote, leave it out on purpose and make sure your contact details are in the approved content so people know where to go.
- If the question is repeated often, move the answer to a clear, dedicated file or website page so it is easy to find.
Keep the content up to date
A bot trained on last season's files will confidently repeat last season's prices. Make updating part of the work you already do. When you change a price list, a policy or opening hours, update the matching source and review the changed parts again before approving them.
A simple routine is enough for most teams: a short weekly look at the unanswered questions, plus a check whenever a price, rule or schedule changes. Put a date inside your documents so you can tell at a glance which version is current, and delete versions you no longer stand behind.
Cost and next steps
Building and testing the bot in the panel is free, with no card needed. When it answers real customers through the widget or the API, each request costs 650 Toman from a prepaid wallet; the minimum top-up is 100,000 Toman, which is about 150 requests. The website widget is a one-time 1,490,000 Toman. There is no subscription and no hidden fees. Details are on the pricing page.
Once your documents are ready, create a free account, add them and the site address, and run your test questions. If you are starting from scratch, the guide on how to build a chatbot for your website covers tone, install and launch, and the docs explain the widget and the API.
Frequently asked questions
Which file types can I use to train the chatbot?
Ravano accepts your website address, PDF, Word, Excel and CSV files, a Telegram channel, Instagram and free text. You can combine several sources, then review and approve what the bot learned before it answers customers.
Will a scanned PDF work?
It may not. A PDF that is only an image of a page often has no text for software to read. Open the file and try to select or search the text. If you cannot, upload the original Word or Excel file, or retype the key parts and add them as free text.
How do I stop the chatbot from giving outdated answers?
Remove old versions of files before you start, put dates on prices and promotions, and review the learned content before approving it. After any change to a price or policy, update that source and approve the changes again. Checking the unanswered questions weekly also helps.
What happens if the answer is not in my documents?
Ravano answers only from information you approved and does not guess, so the bot should say it does not have the answer and point the customer to you. The panel lists the questions it could not answer, and you can add the missing content and approve it.
How much does it cost to train a chatbot on my documents?
Building and testing is free, with no card. Answering real customers costs 650 Toman per request from a prepaid wallet, with a minimum top-up of 100,000 Toman, about 150 requests. The website widget is a one-time 1,490,000 Toman, and there is no subscription.
