Knowledge Base

The knowledge base is the content an agent answers from and the boundary of what it is allowed to say. In a chat or voice agent it is the input to retrieval: documents go in, passages come out at question time, and the model writes from those passages. It is also the thing a support lead actually controls. Prompts shape behavior and tools provide live data, but the knowledge base decides whether the agent knows your returns window is 30 days.

How it works. You upload files (PDF, DOCX, spreadsheets, text) or point the platform at a website to crawl. Each document is split into passages, embedded and indexed. When a customer asks a question, the retriever pulls the closest passages and the model answers from them, declining or handing off when nothing matches. Changing a document re-indexes it; removing one takes its answers out of circulation. The practical inputs are a folder of documents and an owner; the outputs are grounded answers and a log of questions nothing in the base could answer.

Why it matters for consistency and cost. One base feeding every channel means the WhatsApp answer about deposits matches the phone answer. It also moves most agent maintenance out of engineering: a price change is a document edit, not a prompt change or a retraining job. The unanswered-question log is the cheapest product research you will get, because it lists exactly what customers asked that you never wrote down. A neglected base does the opposite: the agent faithfully repeats last year's hours, and every wrong answer looks like an AI failure when it is a content failure.

Example. A gym uploads its membership tiers, class schedule, freeze policy and a page on guest passes. A member asks on web chat whether she can freeze for two months while traveling. The agent retrieves the freeze policy, answers that freezes run one to three months at a reduced monthly fee, states the fee, and offers to submit the request. The next week the schedule changes; the manager replaces the schedule file and the agent's class answers update without anyone touching the agent. A question about childcare hours goes unanswered three times in a week; it shows up in the log, and the manager adds a childcare page.

Knowledge base vs system prompt. The system prompt tells the agent how to behave: tone, what it may never promise, when to transfer. The knowledge base tells it what is true. Facts belong in the base, where they can be versioned and retrieved only when relevant; rules belong in the prompt. Stuffing the price list into the prompt makes every turn more expensive and makes the model more likely to skip a rule.

Common pitfalls.

• Keeping the old and new versions of a policy indexed at the same time, so retrieval returns both.
• Uploading scanned PDFs with no text layer, which index as empty.
• Internal documents with staff-only information that the agent will cheerfully read to customers.
• No owner, so the base decays until someone notices a wrong answer in a review.

Botcadence agents are trained from uploaded documents and website crawls, with the website widget typically live in about 5 minutes after upload; see document training and chat agents.