Give a Knowledge Base to an agent or a prompt

An agent does not know your documents unless you give it a Knowledge Base. You give it with a step named Knowledge Base, which you plug into the agent. The agent then searches the base itself, when it needs to.

• An agent with a model. See Create an agent.
• A Knowledge Base that holds something. See Create a Knowledge Base.

In our example, the Reminder Writer must follow the payment terms of the company, kept in the base Payment terms.

1. Open the agent in its editor.
2. In the Nodes catalog, search for Knowledge Base. It is under AI > Knowledge.
3. Drag Knowledge Base onto the canvas, below the agent.
4. Drag from its diamond to the diamond of the agent named Knowledge Bases. It is the second one from the left.
5. Right-click the step and select Edit.
6. Next to The Knowledge Base it searches., select Change…. The chooser Select a knowledge base opens. Pick Payment terms and select Select. If you have none yet, New knowledge base creates one there, and it is read at once.
7. Select Apply, then Publish.

You can plug several bases into the same diamond, one step for each. The agent then has one search for each base.

The agent receives one search tool for each base. The tool is named after the base: for Payment terms, it is retrieve_payment_terms. The agent decides when to use it. It gives a question, in its own words, and gets back the closest passages, each with its title, its source and a score.

The sentence that you wrote in What it holds is told to the agent with the tool. That is how it knows what the base is for. Write it well. See Create a Knowledge Base.

The agent must be in the execution mode Tools, which is the default. In the mode Conversational, an agent never uses a tool. See What an agent is.

Add a sentence to the instructions of the agent, for example:

Follow the payment terms of your knowledge base. If they do not cover the case, say so.

An LLM Prompt step also has a diamond named Knowledge Bases. Plug the same Knowledge Base step into it.

The base is not searched by the model here. It is searched once, with the text of the prompt, and the passages found are added to the prompt before the model answers. The output knowledge_used of the step gives what was found.

1. In the window of the base, tab Used by: the agent is listed. See Look inside a Knowledge Base.
2. Run the agent with a request whose answer is only in the base. The e-mail must give your payment terms, as written. See Run an agent and read its answer.
3. In the Execution view, the Inspector shows the steps of the execution.

• The agent does not search the base. Check that the step is plugged into the diamond Knowledge Bases, and not into AI Agent Tools. Check the execution mode. Say in the instructions to use its knowledge.
• The agent answers with something else than your documents. Try the same question in the tab Try a search of the base. If the right passage does not come first, change the source.
• The search says that the base holds nothing yet. The first reading failed. Open the base and read What could not be read.

• What the agent finds in a base is sent to the AI model, like the rest of the conversation. Do not put in a base what must not leave your server.
• A Template that carries the agent carries the base too, without its documents. See Share your work as a Template.

Where each kind of information goes, to choose between a Knowledge Base, a Data Table and a session.