# How to Build a Chatbot That Answers from Your Website and Documents

Canonical HTML: https://troymaya.com/chatbot-website-company-documents/
Markdown: https://troymaya.com/chatbot-website-company-documents.md

Published: 2026-10-19T14:00:00+00:00
Author: Troy Ochowicz

![A small stack of approved documents and website cards connected by elegant threads to one clear conversational interface.](https://troymaya.com/wp-content/uploads/2026/10/tm-howto-12-chatbot-company-documents-hero.webp)

To build a chatbot that answers from your website and documents, choose a limited set of approved sources, connect them to a retrieval system, instruct the assistant to use the retrieved evidence, and test questions the sources can and cannot answer. Start with one useful task, such as explaining services or helping visitors find a policy.

This approach is often called retrieval-augmented generation, or RAG. The system looks up relevant material when answering. Uploading documents to a knowledge base is not necessarily retraining the underlying model, and it does not guarantee that every answer will be correct.

## 1. Define a small first version

Use public information for the initial prototype. A practical source set might contain a service overview, business hours, a contact page, and ten approved FAQs. Exclude draft policies, duplicated old files, customer records, and documents the chatbot’s audience should not see.

For a fictional example, Oak Studio wants visitors to understand its design services and submit inquiries. Its assistant may explain the published process but may not promise a project price or delivery date. That boundary makes both the build and the test manageable.

## 2. Prepare an approved source register

```
Source ID: FAQ-01
Title: Public service questions
Owner: Studio manager
Approved version: 2026-10
Audience: Public
Source location: Approved website page or file
Review trigger: Service or policy change
Allowed use: Explain listed services
Excluded use: Binding quotes and availability promises
```

Give each source a clear title and remove contradictory versions from the active collection. If the same answer appears in several places, choose the authoritative version and bring the others into agreement. Keep source ownership outside the chatbot prompt so staff can maintain it.

![Answers from approved sources: Select useful documents; Retrieve relevant passages; Answer with evidence; Escalate missing information.](https://troymaya.com/wp-content/uploads/2026/10/tm-howto-12-chatbot-company-documents-workflow.webp)
Answers from approved sources: Select useful documents; Retrieve relevant passages; Answer with evidence; Escalate missing information.

## 3. Connect sources and inspect what was imported

Choose a platform that supports your document types, source updates, access requirements, and website integration. As one implementation reference, [Voiceflow documents its supported knowledge-base imports](https://www.voiceflow.com/docs/documentation/build/importing-data-sources). Check the current limits for the tool you select rather than assuming every PDF or website page will be read correctly.

After import, inspect extracted text or retrieved passages if the platform exposes them. Test a heading, a number, and an exception from each important document. A successful upload only proves the file was accepted; it does not prove that a table retained its column relationships.

## 4. Configure answers around evidence and limits

A useful starting instruction is: “Answer service questions using the approved sources. If they do not establish the answer, say what is missing and offer the inquiry route. Do not invent prices, availability, or policies.” Adapt this to your workflow and test it. Instructions alone are not an access-control system.

Where supported, display a source link that the visitor can actually open. A source-based answer should distinguish a general policy from a conclusion about the visitor’s specific situation. If the documents explain the quote process, the chatbot can explain that process; it cannot infer the final quote.

## 5. Test supported, missing, and conflicting questions

For Oak Studio, “What services do you list?” should match the approved service page. “Can you finish my project next Friday?” should be referred for review when availability is absent. “An old brochure says something different” should trigger a source-version check rather than a confident blended answer.

Record the question, expected source, actual answer, and outcome. Test paraphrases and follow-up questions, not just the exact wording of an FAQ. Run a deliberately out-of-scope request and a request for information you did not supply. Fix source quality or retrieval before simply adding more prompt instructions.

## 6. Launch with a human route and a maintenance owner

Embed the assistant on a limited set of relevant pages first. Check desktop, mobile, keyboard access, contact links, and what happens when the service is unavailable. Assign someone to review unanswered questions and update sources when the business changes. Treat the initial launch as a monitored workflow, not a finished source of truth.

## Need a document-based assistant?

TroyMaya offers [practical AI assistants built around organizational information](https://troymaya.com/ai-services/). [Discuss your document chatbot](https://troymaya.com/contact/) with your website, document types, intended audience, and examples of questions it should answer. Start with public or redacted material.

Prepared with AI assistance. Examples are illustrative unless stated otherwise.
