How to create a knowledge base is one of the most common questions from teams that have already decided to implement an AI agent and suddenly faced reality: in most companies, the necessary knowledge is scattered across PDF files, Slack messages, the heads of experienced employees, and outdated Excel spreadsheets. Without a quality knowledge base, even the most expensive AI model will hallucinate, give inaccurate answers, or simply not know what to say. In this guide, we explain what knowledge base structure is needed for AI and provide a ready rubricator template.
Data Preparation for AI: Why It Is the Most Important Step
There is a persistent illusion: “We’ll just upload all our documents — and AI will figure it out.” Unfortunately, that is not how it works. Data preparation for AI is the process of transforming unstructured corporate information into a format that can be easily scanned, indexed, and used by the model to generate accurate responses.
Typical problems with poor-quality knowledge bases:
- Contradictory information: the same question described differently in two documents.
- Outdated data: current prices and terms have long since changed, but the base still shows old ones.
- Excessive filler: long corporate reports where the useful answer is buried inside a five-page paragraph.
- Missing structure: a stream of continuous text without headings or sections.
Knowledge Base Structure: The Basic Rubricator
A universal knowledge base structure for customer support is built on the principle from general to specific. Here is a basic rubricator that suits most companies:
Section 1. About the company and product. What we do and for whom – Key products/services with brief descriptions – What is not part of our offering
Section 2. Orders and payment. How to place an order – Accepted payment methods – What to do if payment fails
Section 3. Delivery and timeframes. Delivery timeframes by region – Order tracking – What to do about delays
Section 4. Returns and guarantees. Return conditions – Timeframes and procedure – What cannot be returned
Section 5. Technical support. Common technical problems and solutions – How to reach a technical specialist
Section 6. Promotions and loyalty programs. Current promotions and terms – How the bonus program works
Chatbot Instructions: Writing Rules
Principle 1. One document — one topic
Do not mix delivery information with warranty information in one file. AI finds answers more effectively when the topic is clearly bounded.
Principle 2. Write as if for a first-time responder
Avoid abbreviations without explanation, internal jargon, and references like “see paragraph 3.4.” The agent responds the way the base is written.
Principle 3. Question–Answer format
- Question: How much does delivery to Kyiv cost?
- Answer: Delivery to Kyiv costs $2.50 for orders under $15, and is free for orders over $15. Timeframe: 2–3 business days.
Principle 4. Update regularly
The knowledge base is a living document. Assign a responsible owner and set a review cycle: for example, monthly for prices and terms, quarterly for the overall structure.
Knowledge Management: From Document to AI Response
Knowledge management in the context of AI support is not merely storing files — it is a systemic process: collection → structuring → upload → testing → updating. Companies that build this cycle receive an agent that becomes smarter over time rather than becoming outdated.
The good news: you can start small. Even a basic document with 30–40 Q&A questions is enough to launch an AI agent that covers most typical requests. If you want to see how your knowledge base “comes to life” in a real AI agent — upload your first document in the Intelswift free trial and launch your agent today.
