HomeBlogLeadsWhat to put in the knowledge base so the model stops guessing

What to put in the knowledge base so the model stops guessing

Читать по-русски

Prices, boundaries, frequent questions and prohibitions — with the omissions that cause most of the damage in practice.

What to put in the knowledge base so the model stops guessing

Whatever the model says to a customer comes from what you wrote down, plus invention where you wrote nothing. The knowledge base exists to shrink the second part.

What has to be in it

Frequent questions with your answers

Twenty real questions from real conversations, answered the way you would answer them. This is the most useful section and the fastest to write, because the material already exists in your message history.

Prices and what changes them

Including what is not included. "From X" without saying what is inside produces an argument later.

Boundaries

What you do not do, whom you do not work with, what you do not promise. Without this the model is agreeable to everything.

Prohibitions

A list of things never to say is shorter and more effective than a list of things to say: no guarantees of results, no deadlines without checking, no comparisons naming competitors.

The omissions that cost most

  • No timelines. The model starts promising "quickly", which is a commitment.
  • No payment and refund terms. Asked in almost every conversation.
  • No tone. Defaults to corporate, reads as foreign.
  • No escalation rule. Then it answers questions it should hand over — when to hand over.

Keeping it current

A knowledge base written once and never touched starts lying the first time your prices change. Tie it to the same moment you change anything customer-facing, or it becomes a source of confident wrong answers.

How to check it

Question it adversarially before customers do, then read real transcripts — tuning a model to your product and what not to trust it with.

How Neurogram does this

Everything above is manual work that runs into the number of hours in a day. Neurogram does it on your own accounts and without you: it collects channels and audiences, writes comments and replies with a language model, and keeps limits and quiet hours so the accounts survive. See pricing or look at your case.

Topic: Leads

More on this topic: The tool for this

Read next

Tuning a model to your product and your rules

The model knows language and not your business. Everything it tells a customer comes from your brief — and from guessing where the brief is silent.

When a model writes the opener: where it actually helps

How a generated message differs from a template with substitutions, and what not to expect from it.

What not to trust a model with when it talks to customers

A model answers confidently whether or not it knows. Which categories of message must never be generated, and why.

Ready to put this to work?

Neurogram automates parsing, commenting and DM conversations on Telegram. Three days free — no card required.

Try it free