Skip to main content

AI chatbot for business websites that answers only from your own documents

A visitor asks about delivery, returns or opening hours and gets an answer taken from your price list, terms or manual, with a link to the exact source. If the answer is not in your documents, the chatbot says so and offers a way to reach your team. It does not make things up.

  • Every answer links to its source document
  • The chatbot on webforte.cz runs on our own product, Dotazio
  • Your documents and conversation logs belong to you
  • You talk directly to the lead developer
Discuss a chatbot for your websiteHow the chatbot answers

AI chatbot for business websites in brief

The AI chatbot for business websites from Webforte is a customer support assistant that answers website visitors using only your company's own material: web pages, price lists, terms and conditions, manuals and FAQs. Each answer includes a link to the source it is based on. When the answer is not in the documents, the chatbot says so and offers contact with a person instead of guessing. We agree in advance which topics it may answer, and it passes everything else on. Conversations are logged so you can see what people ask and what is missing from your documents. The work draws on our own product Dotazio, an AI chatbot for Czech municipalities and companies that also runs on webforte.cz. Webforte Technologies s.r.o. (company ID 23364343) is based in Prague, Czech Republic, and works with clients abroad. Pricing is set after an initial consultation.

How our chatbot answers

The chatbot does not treat the language model's general knowledge as a source of facts. The model is there to understand the question and to phrase clearly what it found in your documents.

  1. 1

    Question

    The visitor types a question in their own words, typos and casual phrasing included: “how long does shipping to Ireland take?”

  2. 2

    Search your documents

    The chatbot searches your documents by meaning and by exact terms, then picks the passages that best match the question.

  3. 3

    Answer with a citation

    The answer is built only from those passages and links to the source, such as your Delivery page or a clause in your terms.

  4. 4

    Hand over to a person

    If the answer is not in the documents or the question falls outside the agreed topics, the chatbot says so and offers phone, email or a contact form.

Result: no invented prices, deadlines or promises. Either an answer from a source, or an honest “I don't know that, here's who to ask”.

When a website chatbot does more harm than good

Most disappointing chatbots fail for the same reasons. If you recognise one of these in your current setup or in a proposal you are weighing up, fix it before the chatbot talks to customers.

  • It invents prices and deadlines

    A general AI chat with no link to your documents answers confidently even when it knows nothing. Customers then quote a discount or delivery time you never offered.

  • It gets stuck in a decision tree

    A rule-based chatbot offers buttons, but any question the author did not foresee ends in a loop of “I didn't understand, please choose an option”. Visitors close the chat and send an email.

  • It doesn't know when to hand over

    Complaints, returns disputes and larger enquiries don't belong with a bot. Without a clear handover, you lose exactly the conversations that matter most.

  • Nobody reads what people ask

    Conversations pile up, but nobody checks which answers are missing from the documents. Six months later the chatbot answers as badly as on day one.

  • It slows the website down

    A widget that loads together with the page makes mobile load times worse. A slower site loses visitors before they ever open the chat.

How we build a chatbot that doesn't make things up

Chatbot quality depends less on the choice of model than on the work around it: documents, search, instructions and testing. We have tested this on our own product, Dotazio.

The documents are half the job

The chatbot knows nothing that isn't in its documents, and when documents contradict each other it answers from whichever version it finds first. So before building anything we go through your website, PDFs, price lists, terms and the standard replies your team sends. We look for contradictions (two different delivery times), outdated information and answers that exist only in a salesperson's head, and we write those down with you. Each document gets an owner who keeps it current.

  • Inventory of documents with owners and last update date
  • List of contradictions and outdated passages to fix
  • Answers that were never written down, now documented
  • No personal data or internal prices in the documents

Search that finds the right passage

The customer asks “what does postage to Germany cost?”, the document says “DE shipping rates”. We combine search by meaning with exact-term search so the chatbot finds the right passage even for product codes, clause numbers or service names. Long documents are split into sensible sections along their headings, not by character count, so an answer never includes half of an unrelated chapter. Every section keeps a link to its original page, which later becomes the citation.

  • Search by meaning and by exact terms
  • Documents split by structure, not mechanically
  • Every passage carries a link to its source

Instructions that rule out guessing

The model gets clear rules: answer only from the passages provided, cite the source, and if the passages don't contain the answer, say so and offer a contact. We define which topics the chatbot may answer and which always go to a person, such as complaints or custom quotes. Tone, form of address, answer length and how the chatbot introduces itself follow your company's style. The chatbot always makes clear that it is a chatbot, not a person.

  • Answers only from retrieved passages, with a citation
  • A list of topics that always go to a person
  • Tone and form of address matching your communication
  • The chatbot never pretends to be human

A test set built from real questions

Before launch we build a set of questions from your actual emails, calls and form submissions, including casual and misspelled ones. Each question gets an expected answer and source. We add questions whose answer is deliberately missing from the documents, plus attempts to pull the chatbot off topic or make it reveal its instructions. The set runs again after every change to documents or instructions, so fixing one answer doesn't break another.

  • Questions from your real customer communication
  • Questions with no answer in the documents: the chatbot must admit it
  • Attempts to bypass instructions or drift off topic
  • Re-run after every change

Your website stays fast

We load the chat widget only after the page has rendered, outside the part that decides how quickly content first appears. That is how it runs on webforte.cz, where the Dotazio widget is live and the site scores 100 in Lighthouse. Visitors who don't need the chat aren't kept waiting for it, and the metrics Google uses to judge page speed are not affected.

  • Widget loads after the page has rendered
  • Lighthouse measured before and after launch
  • The same setup runs on our own website

Safeguards against hallucinations and data leaks

No language model can be made error-free. Errors can be contained so that they are rare, traceable and fixable. We set up these safeguards on every chatbot.

Source of answersThe model's general knowledge is not used as a source of facts.
Your documents only
CitationCustomers and your team can check where the information comes from.
Source link with the answer
No answer availableInstead of guessing it offers phone, email or a form, as you choose.
Says so and offers contact
Escalation to a personComplaints, returns disputes, custom quotes and anything else you specify.
Set topics always handed over
Topic limitsIt won't help with homework or comment on your competitors.
No answers outside agreed topics
Conversation logsUnanswered questions show what is missing from your documents.
Used to improve the documents
DataVisitors are told how data is processed, and a retention period is set.
Processed under GDPR, data location agreed with you
Confidential informationThe chatbot can't reveal what it doesn't have, so internal files are filtered out during preparation.
Kept out of the documents

Three kinds of chatbot compared

Three quite different things are sold under the name chatbot. Before comparing proposals, check which kind you are being offered.

Swipe the table sideways to see all columns.

Three kinds of chatbot comparedRule-based chatbotGeneral AI chatOur pickChatbot built on your documents
Where facts come fromPre-written repliesThe model's general knowledgeYour documents and website
Unexpected questionDoesn't understand, shows buttonsAnswers even when it doesn't knowAnswers from a source or hands over
Can you verify the answerReply is fixed in advanceNo sourceCites a specific document
MaintenanceEdit the tree by handYou can't control what the model knowsUpdate the documents

A chatbot built on documents won't answer anything that isn't in them. If your information changes daily and nobody writes it down, its value will be limited.

How the rollout works

We set the timeline and price only after reviewing your documents, because their state affects the scope more than the technology does. We work remotely.

  1. Consultation and topic scope

    We go through what customers ask you most, where it costs you the most time and what must stay with a person. We decide where the chatbot sits on your site and where it sends handovers.

    Output: A list of topics the chatbot answers and topics it hands over.

  2. Document review

    We review your website, documents and standard replies, and mark contradictions, outdated information and gaps that need writing up.

    Output: Document inventory with owners and a list of fixes and additions.

  3. Prototype on your documents

    We build a working version on your real documents and put together a test set from real questions. You try the prototype yourself and show it to the people who answer customers today.

    Output: Working prototype and a test set with expected answers.

  4. Tuning and launch

    Based on the test results we adjust documents and instructions, set up the handover to people, the privacy notice and deferred widget loading. Then we switch the chatbot on.

    Output: Live chatbot on your website with handover in place.

  5. Operation and improvement

    We regularly review conversations, especially questions the chatbot couldn't answer. They turn into a list of documents to add. After every change the test set runs again.

    Output: Overview of unanswered questions and proposed additions.

What you get

  • A chatbot live on your website, connected to your documents
  • Cleaned-up, completed documents your own team can use too
  • A list of topics the chatbot answers and topics it hands to a person
  • A test set with expected answers for checking after every change
  • Access to conversation logs and an overview of unanswered questions
  • Text for your GDPR privacy notice
  • A guide to keeping the documents current, with clear ownership

Where we have proved this

Yours would not be our first chatbot. We run Dotazio as our own product in live production, and the PECKA courses show us how people who rarely use AI actually ask it questions.

Who a chatbot makes sense for

A good fit if

  • You answer the same questions again and again: delivery, returns, opening hours, procedures, terms
  • Your information is written down on your website or in documents, or you are willing to write it down
  • You want customers to get answers in the evening and at weekends, with a person taking over complex cases
  • Someone on your team will review unanswered questions now and then and update the documents

Not a good fit if

  • You want to replace your entire customer service, including complaints and negotiation
  • Your information changes daily and nobody records it: the chatbot only knows what is documented
  • You need a chatbot that advises outside your field or stands in for an expert
  • You want to launch the chatbot and then stop looking after the documents

Frequently asked questions

We don't publish a price list, because the cost depends mainly on the state and volume of your documents, the number of topics, how handover to people works and whether the chatbot connects to other systems. A small firm with a tidy website is a different job from one whose knowledge sits in scattered emails. We quote after a consultation and a review of your documents.

Embedding the widget itself is quick. Most of the time goes into preparing documents and testing against real questions. If your information is already well documented, it goes much faster than if we have to gather it first. We confirm a specific timeline after the document review, once we know how much needs adding or fixing.

It says it couldn't find the answer in the documents and offers contact with a person: phone, email or a form, whichever we set up. The question is added to an overview of unanswered questions, so you can see what is missing and fill the gap. The chatbot doesn't guess or invent an answer.

No language model can be made completely error-free, and we don't claim otherwise. We reduce the risk: answers come only from retrieved passages, each cites its source, the chatbot admits when it doesn't know, and set topics always go to a person. Before launch we test it on real questions, including ones with no answer, and we review conversation logs in operation.

The chatbot answers from the documents you provide. If your website and documents exist in English, German or Czech, it works with that content. We agree during the consultation which languages it should cover for your customers and check each one in the test set with questions in that language.

Where conversations are stored and for how long is agreed up front and written into the data processing agreement. Visitors are told they are talking to a chatbot and how the conversation is handled. The chatbot doesn't ask for personal data on its own, and none goes into the documents. We provide a description of the processing for your privacy policy.

Working with a nearshore team in Prague means EU data rules, a Central European time zone and direct contact with the developer who builds your chatbot, not an account manager. Everything runs remotely. We also run our own chatbot product, Dotazio, which you can try on webforte.cz before we even talk.

You do. The documents we prepare together are yours and can be reused elsewhere, for example as an internal handbook for new staff. You have access to the conversation logs and can export them. If you decide to stop using the chatbot, we hand over the documents and logs.

We handle the technical operation. The content is shared work: you know what has changed in your business, and we see in the conversations what the chatbot can't answer. We regularly send an overview of unanswered questions with suggested additions. When you change prices or terms, updating the document is enough for the chatbot to follow the new version.

See how the chatbot would answer your customers

Send us a link to your website and a few of the questions customers ask you most. In the consultation we go through your documents, show what the chatbot could answer and what it would need to hand to you, and prepare a quote.

Free consultation