Node reference

AI nodes

The AI node is where the reasoning happens. Flux ships two — GPT Model (OpenAI) and Gemini Model (Google) — and they behave the same way on the canvas.

What an AI node does

Given its configuration and whatever is wired into it, an AI node will:

  1. Read its input — usually the user's message or a previous node's output.
  2. Read its attachments — conversation memory, and a knowledge base if one is attached.
  3. Decide whether any of its tools should be called, and call them with parameters it fills in itself.
  4. Produce a final answer as its output.
One AI node with tools and memory attached.

Configuring one

SettingWhat it controls
CredentialWhose API key is used — Flux's default, or one of yours from the credential store
ModelWhich model in that family to call
System promptWho the agent is and how it should behave. The single most important field
InputWhat it is being asked, usually a {{POUT ...}} expression

Writing the system prompt

The prompt is where an agent becomes reliable or unreliable. What tends to work:

  • Say what it is. "You are a support agent for Acme, a company that sells industrial sensors."
  • Say what it must not do. Refunds, pricing commitments, medical or legal advice — name the boundaries explicitly.
  • Say what to do when it does not know. Otherwise a model will guess. "If the knowledge base does not answer it, call the transfer tool."
  • Say how to sound. Length, formality, whether to use the customer's name.
Most "the AI gave a bad answer" problems are prompt problems or tool description problems, not model problems. Before switching models, read the compiled log and check the model actually received what you think it did.

Tools

Tools attach to the node's tool connector, not its input. The model reads each tool's description and decides whether this request needs it. Describe parameters with the DESC keyword so it knows what to fill in.

See Tool nodes.

Attachments

Memory

GPT Memory or Gemini Memory carries the conversation through the session, so the agent can follow "and what about tomorrow?".

Knowledge Base

Your documents, so the agent answers from your material rather than from what the model happens to know.

Use the matching memory for the model — GPT Memory with GPT, Gemini Memory with Gemini.

See Attachment nodes.

GPT or Gemini?

Both are first-class in Flux, take tools and memory the same way, and are wired identically. The practical differences are cost, speed and which provider you already have keys and an agreement with.

Because they are interchangeable on the canvas, the honest approach is to build with one, keep the prompt, and try the other on the same workflow. Compare the answers and the token cost in usage analytics.

What AI nodes cost

AI nodes are billed in tokens — input and output. A node with several tools and a long memory sends more context on every call, so cost per session grows with conversation length rather than with node count.

Using Flux's built-in keys bills through your project. Using your own credential bills you directly with the provider.