Build a knowledge base your AI agent can use

When an AI agent gets an answer wrong, the cause is usually in the documents, not the model. Out-of-date pricing, two versions of a policy and a page that only makes sense to the person who wrote it all produce confident, wrong replies. The fix is editorial work you can do before launch.

1. Start from the questions, not the files

Export the last few months of support tickets, chat logs and sales emails. Group them by topic and count them. The top 50 questions are your launch scope. Anything outside that list can wait.

2. Pick one source of truth per topic

• Name the document that owns each policy: refunds, shipping, pricing, eligibility, service levels.
• Remove or archive older copies. If two files disagree, the agent has no way to know which is right.
• Date every policy and put the date at the top of the file.

3. Write for the question

A document written for a lawyer is hard for anyone to retrieve from. A short section that begins with the question and answers it directly works far better.

• Use headings that read like questions: "How long do refunds take?"
• Answer in the first sentence, then add the detail and the exceptions.
• Keep one topic per section so the agent can quote it cleanly.
• Spell out terms, plan names and product codes the way customers say them.
• Add a short "what we do not do" page. Honest limits prevent made-up answers.

4. Choose the right size for your plan

Avarole's AI knowledge base answers from the documents you upload and cites the source file, so a reader can check it. Storage and retrieval depth scale with the plan.

Plain text, FAQs and policies are small, so most teams are far from the limit. Large manuals, contract archives and product catalogues are what move you up a tier. Compare plans on the pricing page.

5. Test before you launch

1. Take 50 real questions from your list, including awkward ones with typos and mixed languages.
2. Ask each one to the agent and record the answer and the source it cited.
3. Mark each as correct, partly correct or wrong. Fix the document, not the question.
4. Add the questions the agent should refuse, such as legal advice or things you do not offer, and check it declines.

6. Review transcripts every week

Transcripts are included on every plan. Read a sample each week and look for three signals: questions the agent could not answer, answers customers pushed back on and topics that keep coming up. Each one is a document to write or correct. After a month the base reflects what your customers actually ask.