RAG explained simply: how a chatbot answers from your knowledge base
What retrieval-augmented generation is, how to build a chatbot on your own data, and how to keep its knowledge base current.
RAG explained simply: a chatbot with RAG looks things up before it answers. For every question, it retrieves the matching passages from your knowledge base and writes its answer from them, instead of relying only on what the language model learned in training. RAG stands for retrieval-augmented generation: text generation backed by a lookup step.
Why a chatbot needs a knowledge base
A language model on its own doesn’t know your prices, opening hours or processes. Asked about them, it sometimes fills the gap with a plausible-sounding invention, known as a hallucination. With RAG, the relevant passages are in front of the model before it answers. Think of it as an open-book exam: the question is no longer “What does the model know about this?” but “What does your content say about this?”
That has two practical consequences. First, the model doesn’t need to be trained on your data: when your content changes, it is ingested again, and the next answer draws on the new version. Second, you can show where an answer came from.
The two steps in plain terms
- Look it up: The system searches your knowledge base for the passages that fit the question. Kyros combines search by meaning with keyword search, then ranks the results again by relevance (reranking).
- Answer: The best passages go to the language model along with the question, and the model writes the answer from them. When the answer draws on your pages, Kyros lists numbered source links to the original page below it, under “Further Links.”
For how chunking, embeddings and reranking work in detail, and what determines search quality, see the glossary entry RAG chatbot.
What RAG can’t do
RAG lowers the risk of wrong answers but doesn’t remove it. Retrieval can surface the wrong passage, and the model can misstate a correct one. When your content has nothing that fits, the assistant in Kyros is instructed to say so. That is an instruction to the model, not a guarantee. Above all, RAG is only as good as the knowledge base behind it: an outdated page produces an outdated answer, complete with a source link.
What belongs in your chatbot’s knowledge base
A chatbot built on your own data can only answer as well as its content allows. Four principles help.
Pages with concrete answers
Add pages that answer common questions directly, such as FAQs, shipping terms, opening hours or service descriptions. General marketing copy rarely helps the assistant.
One current version per topic
If the same deadline appears differently on two pages, retrieval may find the outdated one. List only the current page. Kyros detects duplicate pages.
Your customers’ words
Name products, plans and forms the way customers ask about them, and put conditions in the same paragraph as the rule. That makes it more likely that rule and condition end up in the same passage the search finds.
Golden answers for exact wording
For questions where every word counts, such as prices, deadlines or return rules, write a Golden answer. When a question matches it, the assistant is instructed to use it word for word.
How to build a chatbot on your own data
The process in Kyros, from the first page to correcting individual answers.
- List your pages: Create a knowledge base and paste the URLs of the pages the assistant should know, one per line. Kyros fetches the pages, splits them into passages and makes them searchable; duplicate pages are detected. Kyros doesn’t crawl a whole site or sitemap, and you can’t upload files such as PDFs.
- Set up re-sync: For each page, choose whether Kyros re-syncs it every 6, 12 or 24 hours or weekly. Changes to your pages reach the assistant with the next re-sync.
- Write Golden answers: For questions where exact wording matters, add the question and answer yourself. Short, unambiguous answers work best.
- Test, publish, embed: Assign the knowledge base to your assistant and test it in the Studio preview. Changes to its settings stay in a draft until you click Publish. Then embed the chat widget with a code snippet; there are install guides for HTML, React/Next.js, WordPress, TYPO3 and the Government Site Builder (GSB). Alternatively, your application can query the assistant through the REST API.
- Close gaps: Chatlogs show conversations and visitor ratings. Where an answer was wrong or incomplete, Revise saves a better version as a Golden answer; if a topic is missing entirely, add a page that covers it.
When the knowledge base has no answer
Kyros has no live chat and doesn’t hand conversations over to staff. Instead, an Interactive Card (Beta) can collect a contact request right in the chat, for example with name, email address and topic. Submissions are stored per Card and can be exported as CSV.
RAG and knowledge base FAQ
What is RAG, explained simply?
RAG stands for retrieval-augmented generation. A chatbot with RAG first searches a defined knowledge base, such as the pages of your website, for passages that match the question, then writes its answer from them. In other words, it looks things up instead of drawing only on the language model’s training. That makes answers about your prices, deadlines or processes possible, and the answer can show where it came from. RAG still doesn’t make a chatbot error-free.
What is a chatbot knowledge base?
A knowledge base is the collection of content a chatbot draws its answers from. In Kyros, it consists of web pages whose URLs you list and Golden answers you write yourself. The pages are split into passages and made searchable. A knowledge base belongs to the workspace and can be assigned to several assistants, so you maintain shared content only once.
How do I build a chatbot on my own data?
With Kyros, you create a free account and add an assistant. Then you create a knowledge base, assign it to the assistant and paste the URLs of the pages it should know, one per line, and write Golden answers for questions where exact wording matters. You can’t upload files such as PDFs. In Studio, you test the draft in the preview and take it live with Publish. Then you embed the chat widget with a code snippet or use the REST API.
What content works best in a knowledge base?
Pages that answer common questions directly: FAQs, shipping and returns policies, opening hours, service and pricing pages, or help articles. Each topic should exist in only one current version so the details don’t contradict each other. Kyros uses only the pages whose URLs you list; it doesn’t crawl a whole site or sitemap, and you can’t upload PDF, Word or Excel files. Put that content on a web page or add it as a Golden answer.
How do I keep the answers up to date?
By maintaining the knowledge base. For each page, you can have Kyros re-sync it every 6, 12 or 24 hours or weekly; changed content counts from the next re-sync. Update Golden answers whenever prices or deadlines change, for example. Chatlogs show conversations and visitor ratings, and Revise saves a better answer as a Golden answer. You make these corrections yourself.
Does RAG prevent wrong answers completely?
No. RAG lowers the risk because the model works from matching passages, but retrieval can surface the wrong passage and the model can misstate a correct one. When nothing fits, the assistant in Kyros is instructed to say so; that is an instruction, not a guarantee. That’s why every widget permanently notes that answers are AI-generated and may be inaccurate, and numbered source links below the answer make checking easier.
Further reading
RAG chatbot
How it works, what drives quality and where the limits are
Knowledge base
Add pages, re-sync them automatically and assign them to several assistants
Golden answer
Fixed answers for questions where the exact wording matters
Create an AI chatbot
Step by step to your own assistant on your website
Build a chatbot that answers from your knowledge base.
The Starter plan is €0, permanently, with no credit card required. It includes one assistant with a chat widget.