How HighLevel Conversation AI uses your knowledge base

A knowledge base and a bot are two separate things in HighLevel. You build one, you attach it to the other, and only then can the bot draw on it. The problems below live in the gap between those steps.

Where the bot gets its answers

Conversation AI retrieves information from its connected knowledge base. In the refreshed interface, the sources themselves are managed under AI Agents > Knowledge Base, and each bot's Bot Training settings choose which knowledge base it uses.

The two are stored apart. HighLevel's agent API returns the knowledge base ids in their own field, separate from the prompt text. That is the right way to picture the build: the facts sit in a knowledge base attached to the agent, not pasted into the prompt.

In our builds that shapes the prompt as well. We keep it short, around a kilobyte, and limit it to what changes from one business to the next: who the bot is, its goal and the fields it collects. That size is our habit, not a HighLevel limit.

Attach the knowledge base to the bot. Don't paste it into the prompt.

Why the bot isn't using it

When answers ignore content you know is there, look for one of these causes:

• It isn't selected. The knowledge base exists, but this bot's Bot Training points at a different one, or none.
• It hasn't finished processing. New content isn't used until processing succeeds and the attachment is complete.
• Bot Training is hidden. With the user's training permission off, Bot Training doesn't appear. The same permission controls attaching and detaching knowledge bases, and it affects whether sources show in a reply's details.
• An old version is still attached. A replaced source can sit beside the new one, and the bot can draw on either.
• A trigger is choosing other content. With Smart Triggers set up, what the bot reaches for depends on their conditions.

Connect it and check it, in order

Work from the permission down to the reply, so each check rests on the one before it.

1. If Bot Training is missing, ask an admin to turn on View & Manage Conversation AI agent Training for your user.
2. Open the bot, go to Bot Training and confirm the knowledge base you expect is selected.
3. Wait for its sources to finish processing successfully before you test anything.
4. Ask the Retrieval Tester a question a customer would ask. It shows whether the content can be found at all.
5. Test the whole bot with a fresh scenario. That also exercises the instructions, goals and actions, which the Retrieval Tester doesn't.
6. Open Response Info on a reply to see which knowledge base sources it drew on, and remove any outdated version you find there.
7. If you use triggers, review the Always-On content and each trigger's conditions.

Always-On and Smart Triggers

Knowledge base triggers decide which content the bot looks at. Smart Triggers select content when their conditions match. Always-On supplies general knowledge when no Smart Trigger matches. New ones are added with Add Trigger.

Keep trigger conditions from overlapping. When two triggers can match the same message, it gets harder to predict which content the bot will use, and harder to test.

HighLevel's pages disagree on how many knowledge bases one trigger can hold: the help article gives one figure, and the API reference gives different figures for different modes. HighLevel has changed this before, so check the current screen before planning around a number.

Answer, then steer back to the goal

A knowledge base answers questions. It doesn't move the conversation on. The rule we write into our chat prompts is answer, pivot, ask: answer the customer's question from the knowledge base, connect it back to the bot's goal, and ask the next question, all in the same message.

That is a prompt convention of ours, not something HighLevel does on its own. Test it the way you test retrieval, in a full conversation, and read the replies as a customer would.