How to Add an AI Chatbot to Your Website That Answers From Your Content

A large language model (LLM) such as GPT, Claude or Gemini knows a lot about the world. It does not know your prices, your returns policy or the opening hours of your Leeds branch. If you ask it anyway, it may guess. Those confident guesses are often called hallucinations.

RAG fixes this by adding a search step before the model answers:

1. Retrieve: the visitor's question is used to search your own content for the most relevant passages.
2. Augment: those passages are added to the prompt, with an instruction such as "answer only from the sources below".
3. Generate: the model writes a natural reply based on those passages, ideally with links to the source pages.

Think of it as an open-book exam. The model is good at reading and writing. RAG hands it the right pages of your book before it answers.

Why not just train a model on my content?

Fine-tuning changes how a model writes. It is a poor way to teach it facts that change, such as stock levels or policies. RAG is cheaper to run, easier to update (you just re-index a page) and lets you show which source each answer came from. For most business websites, RAG is the right starting point.

Gather these before choosing any tool. It saves time later.

• Your source content: website pages, FAQs, product or service sheets, policies and help articles. PDFs and Word files work too.
• A clear scope: what the bot should answer (for example, services, pricing approach, delivery, support) and what it must not (legal advice, account-specific data).
• A handover route: an email address, contact form or live chat for questions the bot cannot answer.
• A list of 30 to 50 real questions from your inbox, sales calls or support tickets. You will use these for testing.
• Someone who owns the content and will update it when things change.

There are three common routes. The right one depends on your budget, your data and how much control you need.

A custom build typically uses an embedding model, a vector database (such as PostgreSQL with pgvector, Qdrant or Pinecone) and a framework like LangChain or LlamaIndex, or plain code. If you need it connected to your CRM or booking system, our AI chatbot development team usually takes this route.

1. Audit and clean your content. Remove outdated pages, duplicate FAQs and conflicting policies. The bot will repeat whatever you give it, so contradictions become wrong answers.
2. Export the content. Use your sitemap (usually at /sitemap.xml) to list pages, or export from your CMS. Convert PDFs to text and keep the page URL with every document so answers can link back.
3. Split it into chunks. Break each page into passages of a few short paragraphs, split at headings where possible. Add a small overlap between chunks so sentences are not cut off from their context. Store the title and URL with every chunk.
4. Create embeddings. An embedding model turns each chunk into a list of numbers that captures its meaning. Questions with similar meaning end up close together, even if they use different words ("refund" and "money back").
5. Store them in a vector index. Load the embeddings and their text, title and URL into your vector database or platform. Many teams also add keyword search alongside it (hybrid search), which helps with product codes and names.
6. Write the system prompt. Tell the model who it is, to answer only from the supplied sources, to say "I don't know" when the sources do not cover the question, to cite the source page, and to offer the contact route for anything else. Keep the tone consistent with your brand.
7. Connect retrieval to the model. For each question, fetch the top few matching chunks, pass them to the model with the system prompt, and return the answer with source links. Keep the API key on your server, never in the browser.
8. Add the chat widget to your site. Most platforms give you a short script tag to paste before the closing </body> tag. On WordPress you can add it through your theme's footer settings or a header-and-footer scripts plugin. Load it asynchronously so it does not slow the page.
9. Test with your real questions. Run your list of 30 to 50 questions. Mark each answer as correct, partly correct, wrong or "should have said I don't know". Fix gaps by improving the content, not just the prompt.
10. Launch, monitor and update. Review conversation logs weekly at first. Unanswered questions show you what content is missing. Re-index pages whenever you publish or change them; automate this if you can.

No chatbot is perfect, but these habits reduce wrong answers a lot:

• Ground every answer. Instruct the model to use only the retrieved passages and to refuse politely when they are not enough.
• Show sources. Linking to the page lets visitors check, and makes errors easy for you to spot.
• Keep scope narrow. A bot that handles services, delivery and support well is better than one that tries to answer everything.
• Handle sensitive topics separately. Route pricing quotes, complaints and anything legal or medical to a person.
• Set a low "creativity" setting. Most APIs have a temperature parameter; a low value keeps answers factual.

A chatbot is a new way for data to enter and leave your business, so treat it like any other system.

• Data protection: update your privacy notice. Depending on where your visitors are, laws such as UK GDPR, EU GDPR, India's Digital Personal Data Protection Act, the UAE's PDPL or Saudi Arabia's PDPL may apply. Check where your provider stores and processes data and whether it uses your data for training.
• Do not index private data unless you add login and permission checks. Anything in the index can, in principle, end up in an answer.
• Prompt injection: visitors may try to make the bot ignore its rules. Keep secrets out of the prompt, limit what tools the bot can call, and rate-limit requests.
• Accessibility: the chat window should be usable by keyboard, readable by screen readers and meet WCAG 2.2 AA contrast. Do not let it cover key content on mobile.

If your bot will touch customer records, a short security review before launch is worth the time.

A simple bot on a platform using public website content can go live in days. A custom RAG system with integrations usually takes a few weeks. The main cost drivers are:

• How much content you have and how messy it is.
• Whether you need integrations (CRM, helpdesk, booking, order status).
• The model you choose and how many conversations you expect, since most providers charge per use.
• Hosting for the vector database and the backend.
• Ongoing content updates, monitoring and improvements.

We design and build RAG chatbots that answer from your own content, link to sources and hand over to your team when needed. If you want help choosing an approach or building one end to end, see our AI development services or get in touch for a straightforward conversation.