Glossary
Knowledge base and RAG
A knowledge base holds the information an AI assistant is allowed to answer from; RAG is the method of retrieving matching passages from it before each reply.
RAG stands for retrieval-augmented generation. Instead of relying on the general knowledge of the language model, the system looks up the matching passages in the stored documents for every question and gives them to the model as a basis. That way the assistant answers with the information of the business and not with guesses.
Quality depends on the content. Short, unambiguous entries work well: opening hours, services, prices, directions, procedures. Outdated pages, contradictory details and long running text work badly. It matters who maintains the content and how changes are approved. With small amounts of knowledge, the entire content can also sit directly in the prompt; retrieval only pays off at a larger scale.
In practice
Building one starts with the questions that are really asked on the phone. Collect them for a week, or go through old notes, and write a short, unambiguous answer to each. Every piece of information should exist in one place only, so that no two versions circulate after a change. Name one person who enters changes, for example new prices or holiday closures. Afterwards check with a test call whether the assistant actually gives the new detail, and remove outdated documents instead of adding to them.
An example
A pharmacy stores its opening hours, delivery service, emergency rota and the procedure for ordering prescriptions in advance. An old flyer with last year’s Saturday hours is still on the website. A customer asks how long the pharmacy is open on Saturday and is told the old time, because the system found both versions. Since the flyer was removed and the opening hours are kept as a short entry of their own, the assistant reliably gives the current time.
A common misunderstanding
More is not better here. Uploading every existing PDF, price list and old circular does not produce cleverer answers, only more contradictions. A second misunderstanding is that RAG means the model learns. It learns nothing permanently; it looks things up before each answer. What is not in the knowledge base, or is not found there, the assistant does not know, even if it was mentioned in a call yesterday.
At Call0
During setup, Call0 reads your website and prepares the assistant’s knowledge from it. Answers come from approved knowledge.
Related terms:
Read on:
On this topic:
As of: 12 October 2026