How to train an AI chatbot on your own data (without code)

“Train a chatbot on your own data” is one of the most searched phrases in AI customer support, and one of the most misunderstood. You don't need a data scientist, a GPU or a pile of training examples. Modern support chatbots don't learn your content the way a model learns language. They look it up, every time a customer asks a question. This guide explains how that works, and how to prepare your content so the answers are accurate.

What “training” really means: retrieval, not retraining

There are two ways to make an AI model answer from your information:

• Fine-tuning changes the model itself, by training it further on examples. It is slow, expensive, has to be repeated whenever your information changes, and the model still can't tell you where an answer came from. It's the right tool for teaching a model a style or a format, not facts that change.
• Retrieval-augmented generation (RAG) leaves the model alone. When a question comes in, the system searches your content for the most relevant passages and hands them to the model with the question. The model writes an answer from those passages and cites them.

For customer support, retrieval wins on every count: answers reflect your content the moment you update it, every answer can show its sources, and nothing about your business ends up inside a model. When people say they “trained a chatbot on their data”, retrieval is almost always what they did. It's how Aidfolk works.

How retrieval-augmented generation works, step by step

1. Collect. The chatbot reads your sources: web pages, documents, notes, even records from an API.
2. Split. Long documents are cut into short, overlapping passages of a few hundred words, so a question can be matched with the exact paragraph that answers it rather than a whole manual.
3. Index. Each passage is turned into an embedding, a list of numbers that captures its meaning, and stored alongside its words.
4. Search. When a customer asks something, the question is matched against the passages by meaning and by keywords at the same time. Meaning catches “I can't get in” for a page about passwords; keywords catch exact product names and error codes.
5. Answer. The best passages go to the language model with instructions: answer only from these sources, cite them, and say so when they don't contain the answer.

Aidfolk does all of this for you. You add your content; it handles splitting, indexing, searching and citing. If you're a developer weighing up building it yourself, our guide to building a RAG chatbot goes through each part.

Step 1: Choose the right sources

Start with the content your customers already read, or should:

• Your help center or knowledge base
• Product documentation, setup guides and API references
• Pricing, plans, billing and refund policies
• FAQs and the replies your team sends most often
• Manuals and handbooks as files (Aidfolk reads PDF, DOCX, PPTX, XLSX, HTML, MD, TXT, CSV, ODT, RTF, EPUB)

In Aidfolk, a website is read by following its links or its sitemap, and you choose which pages to keep. Files are uploaded or imported from a link, notes are typed in, and an API source turns records from your own systems (JSON or XML) into passages.

Leave out anything a customer shouldn't see:

• Internal notes, customer data, passwords and keys
• Outdated pages that contradict current ones
• Marketing pages that say a lot but answer little

Step 2: Prepare your content so it's easy to find

Retrieval is only as good as the passages it can find. A few habits make a big difference, and they help human readers too:

• One topic per page or section, under a heading that says what it is (“Reset your password”, not “Account stuff”).
• Put the answer first, then the detail. A passage that starts with the answer is a passage that gets used.
• Use the words your customers use. If they say “invoice” and your docs say “billing document”, mention both.
• Write things down that only live in people's heads. If your team answers a question by memory, it belongs in a note.
• Make sure text is real text. Pages that only appear after JavaScript runs, and scanned PDFs without a text layer, can't be read. Export a text version instead.

Step 3: Give the agent clear instructions

Your content says what is true; instructions say how to answer. Good instructions are short and specific:

• Who it is: “You are the support assistant for Acme Cloud.”
• Tone: “Friendly and brief. Use numbered steps for instructions.”
• Scope: “Only answer questions about Acme Cloud.”
• The way out: “If you don't know, say so and suggest emailing support@acme.example.” In Aidfolk this has its own setting, the reply it gives when the sources don't hold the answer.

Aidfolk tells every agent to answer only from your sources, to cite them, to treat what's written inside your documents as information rather than instructions, and to reply in the language of the question.

Step 4: Test it before customers see it

Collect 20 to 50 real questions from your inbox, including awkward ones, and ask them in the agent's test chat before you put it on your site. For each answer, check three things:

1. Is it right? Compare it with what your best support person would say.
2. Are the sources the right ones? Aidfolk lists them under every answer, with page numbers for PDFs.
3. Does it admit what it doesn't know? Ask about something your content doesn't cover. The right answer is “I don't know”, not a guess.

When an answer is wrong, fix the content first, not the instructions. A missing or unclear page is the usual cause, and fixing it helps every future answer.

Step 5: Keep it current

• Read pages again after you change them. In Aidfolk, “Read again” updates one page and “Read all again” a whole site. The old text keeps answering until the new text replaces it.
• Review unanswered questions weekly. Conversations can be filtered to those with unanswered questions, and Analytics shows the questions customers ask most. Each one is a page you haven't written yet.
• Replace files rather than adding new copies. Aidfolk keeps earlier versions of a replaced file, so you can go back if needed.

What about privacy?

Your content is used to answer your customers, nothing else. In Aidfolk each workspace's passages are kept apart by the database itself, the AI provider receives only the question and the passages needed to answer it, and you decide what goes in. Read more on our security page.

Ready to put it on your site? Follow how to add an AI chatbot to your website.