Definition
The information an AI agent draws on to answer: your website pages, documents, and written questions and answers. The agent searches it before each reply, so its answers come from your facts rather than guesses.
What the agent knows, and where it comes from
A language model knows how to write. It doesn’t know your prices, your opening hours or whether the 12-week program includes nutrition coaching. A knowledge base is where that knowledge lives for an AI sales agent: the pages of your website, your brochures and price lists, and the questions your team answers every week.
When a lead asks something, the agent first searches the knowledge base for the passages that match, then writes its answer from them. The technique is often called retrieval-augmented generation, or RAG. It keeps answers specific, and you change what the agent says by changing a document, without retraining anything.
Building a good one
- Start from real questions. Read your last fifty conversations and list what people actually ask.
- One fact, one place. If the price is in three documents, two will soon be out of date.
- Write the answers you’d give. “Parking is free after 6 pm in the lot behind the gym” beats “parking available”.
- Say what you don’t do. “We don’t offer financing” prevents an invented yes.
- Keep it current. A price change that never reaches the knowledge base becomes a promise you didn’t make.
When answers go wrong
Many bad answers come from the knowledge base, not the model: a missing fact, an old PDF, two pages that contradict each other. The fix is to find which passage the agent used and correct it at the source, rather than adding rules on top.
Knowledge bases in WhatSetter
WhatSetter writes a first knowledge base when you create the agent, from your site, a PDF or a short description. You can then add your site’s pages, re-read every week, your files (PDF, Word, text or CSV, up to 50 MB) and up to 200 questions and answers. In the Test tab, the “Why this reply?” panel shows the criteria and excerpts the agent used, so you can see where an answer came from and fix it. How many knowledge bases, sites and files you get depends on your plan. See the AI sales agent.