What does “training a chatbot on your own data” really mean?
When people say they want to train a chatbot on their data, they usually imagine the AI studying their documents until it knows them by heart. That is not how business chatbots work, and it would be a poor design: a model that memorised your price list would keep quoting old prices after you changed them.
Instead, the chatbot keeps your content in a searchable library. When a customer asks something, it first searches that library for the passages that match the question, then writes a reply from them. The technical name for this is retrieval-augmented generation, or RAG, which simply means “look it up first, then answer.” Changing an answer is a matter of editing the item and syncing it again.
Which kinds of data can you train a chatbot on?
Different content needs different shapes. A price is a number to compare, an opening hour is tied to a branch, and a warranty policy is a paragraph to quote. Gydr keeps four kinds of knowledge base so each is searched the right way.
| Knowledge type | Put in it | Customers ask things like |
|---|---|---|
| FAQ | Questions and answers in your own words | “Do you deliver to Penang?” |
| Product catalogue | Products with name, description, price, link, image and your own attributes such as size or colour | “Do you have a waterproof bag under a certain price?” |
| Documents | PDFs, Word files, spreadsheets, text, web pages and images (PDF, DOCX, XLSX, TXT, MD, HTML, CSV, JSON, PNG, JPG, WEBP) | “What does the warranty cover?” |
| Locations | Branches with address, map position and opening hours | “Which outlet near me is open now?” |
How to train a chatbot on your data in Gydr
These steps use the names on screen in the Gydr dashboard. You can repeat them for each kind of knowledge, and attach several knowledge bases to one chatbot.
- 1Open Knowledge Base, click Create Knowledge Base, give it a name and pick its type.
- 2Add your items: type them in, use the CSV import (download the template, fill it, upload it), upload documents, or choose Import from Website.
- 3Review what came in. Items brought in by a website import start inactive, so you can check them and then use Activate all.
- 4Click Sync Embeddings. This turns each item into a searchable form; an item marked Out of Sync has changed since the last sync.
- 5Open Skills, edit the Skill your chatbot uses, and connect the knowledge base on its Knowledge tab.
- 6Open your widget, click Test widget and ask the questions the content should answer.
How do you train a chatbot on PDFs and documents?
Upload the file to a documents knowledge base. Gydr reads the text, splits long files into passages so the right part can be found, and keeps headings and tables as readable as it can. Scanned PDFs that are only pictures of pages are read with text recognition, which takes longer and uses more credits than a PDF with real text.
Long documents are fine, but clarity helps. A ten-page manual where each section has a clear heading answers better than the same text pasted as one block. If a document mixes old and new versions of a policy, remove the old one: the chatbot cannot tell which one you meant.
How do you train a chatbot on your website?
Use Import from Website on a knowledge base. You can read the whole site from a start page, pick pages from your sitemap, or paste the exact pages you want. Gydr opens the pages in a real browser and pulls out FAQs, products, locations or document text, depending on the knowledge base type.
Before the import starts, you see the pages it will read, and you are charged only for pages it actually reads. Afterwards, review the results before you activate them. Websites carry menus, cookie notices and expired offers, and a quick review keeps those out of your answers.
Why does the chatbot answer only from your content?
A general chatbot such as a public AI assistant answers from everything it read on the internet. For a business that is a risk: it may quote a price you never set or promise a service you do not offer. Gydr is built the other way round.
- Every reply is written from the items found in your knowledge for that question.
- Before a reply goes out, it is checked against those items; a claim the items do not support is corrected or removed.
- If nothing in your knowledge answers the question, the chatbot says it does not have that information instead of guessing.
- A product with no description is presented by name and price only, so the chatbot does not invent details for it.
- Each chatbot searches only the knowledge bases attached to its own Skills, so two chatbots in one account can stay separate.
How do you keep a trained chatbot up to date?
Treat the knowledge base like your price list: someone owns it, and it changes when the business changes. Edit an item, then sync it, and the next answer uses the new content. To hide an item for a while, set it inactive instead of deleting it.
Every week, look at conversations where the chatbot could not find an answer. Each one is either a missing FAQ, a product without enough detail, or a question you have decided not to answer. Adding the missing pieces is how the chatbot gets better over time.
Running a Shopify or WooCommerce store? Instead of copying products in, you can connect the Shopify Storefront or WooCommerce Storefront plugin, and the chatbot searches your live store.
Common questions
You can build knowledge bases on Gydr's Free plan. Syncing, document reading and website imports use credits, and the Free plan includes starter credits to begin with.
No. Edit the item and sync it. There is no model to retrain; the chatbot searches your updated content on the next question.
Yes. Pages that are only images are read with text recognition. Clear scans give better results, and a PDF with real text is faster and uses fewer credits.
Yes. Connect the Shopify Storefront or WooCommerce Storefront plugin to your Skill and the chatbot searches your live store, so stock and prices come from the shop itself.
Less than most people expect. Clear answers to your 20 to 50 most common questions, plus your product list, cover most conversations. Add more as the conversations show you what is missing.