Preparing a knowledge base for an AI assistant

Step 1: collect sources

• Public website content: services, products, prices, delivery, returns.
• Policies and terms.
• Product catalogs and specifications.
• Internal procedures, if the assistant serves staff.
• Existing FAQs and support templates.
• Frequent questions and answers from chats, emails and call notes.

Prefer structured data from systems (prices, stock) over copies in documents, so the assistant reads current values.

Step 2: remove contradictions

Different versions of the same information are the main source of wrong answers. For each topic, identify the authoritative source and remove or archive others. Where information changes often — prices, schedules — connect the assistant to the system that holds it rather than to a document.

Step 3: write for retrieval

The assistant finds passages relevant to a question and answers from them. Content works best when:

• each passage answers one topic and makes sense on its own;
• headings describe the content ("Delivery to regions outside Baku");
• important conditions are stated explicitly, not implied by context elsewhere;
• dates and validity are included for time-sensitive information;
• tables are clean and labelled.

See RAG.

Step 4: fill the gaps

Review real customer questions and check whether the knowledge base answers them. Common gaps: edge cases, exceptions, "what if" questions, practical details like parking or documents needed. Write short answers for them, approved by the responsible person.

Step 5: define boundaries

• Topics the assistant must not answer: legal or medical advice, individual pricing, complaints.
• What to do when the answer is not in the sources: say so and offer a person.
• Information that must not be disclosed: internal margins, personal data, confidential processes.

See AI hallucinations and business risk.

Step 6: languages

If customers ask in Azerbaijani, Russian and English, the assistant can answer in the question's language even if sources exist in one language — but quality is better when key content exists in each language, especially terminology and legal wording. Test in each language. See AI assistants in Azerbaijani.

Step 7: maintain it

• Each source has an owner who updates it.
• Changes in prices, policies or services trigger updates to the knowledge base.
• Unanswered or poorly answered questions from conversation logs are reviewed regularly and turned into new content.

Personal data

Do not include customers' personal data in a knowledge base for a public assistant. For assistants that access customer-specific information, such as order status, use secure lookups with authentication instead of documents.