Introduction
A model is the AI engine your automations use. The AI Kit does not run AI itself — it connects to providers (OpenAI, Anthropic, Mistral) or to a self-hosted alternative (Ollama). Each connection is a model in your workspace.
This section explains what kinds of models exist, what to look for when picking one, and how to connect each provider step by step.
Why models matter
Almost every automation that does something useful with AI calls a model at least once. The model you pick determines:
- What the AI can do. A small model handles short summaries well; a large model can do complex reasoning over thousands of words.
- How fast it responds. Smaller models are typically faster.
- How much it costs. Provider pricing varies by model, by token, and sometimes by region.
- Where the data goes. External providers send the prompt to their infrastructure. Self-hosted models never leave your premises.
You can connect more than one model to a workspace and use different models in different automations — or even in different steps of the same automation.
Two kinds of models
| Kind | Used for | Where you select it |
|---|---|---|
| Chat models | Generating text, answering questions, calling tools, reasoning. | Steps like LLM Prompt, AI Agent, and inside Agents. |
| Embedding models | Turning text into numbers a knowledge base can search. | In every Knowledge base configuration. |
Most providers offer both kinds. When connecting a provider, you can choose to expose only the chat side, only the embedding side, or both. Many automations work fine with a single chat model and a single embedding model.
How to choose
Start with one chat model and one embedding model per provider. Add more only when you have a specific need.
For the chat side:
- Need fast, cheap, good-enough answers? Pick a small model (GPT-4o mini, Claude Haiku, Mistral Small).
- Need quality answers, reasoning, complex instructions? Pick a large model (GPT-4o, Claude Sonnet, Mistral Large).
- Need to keep data on your own servers? Self-host an open-source model with Ollama.
For the embedding side, the choice matters less day to day; what matters is consistency. Use the same embedding model for all knowledge bases you want to search together.
Where models appear in the user interface
The Models tab in the main navigation lists every model connected to the workspace.
Click a card to edit a model. Use Configure new Model to add one. See Create a New Model.
What to do next
- Create a New Model — the generic step-by-step.
- Bring-your-own-Key (BYOK) — the licensing and credentials model.
- Provider-specific pages: OpenAI · Anthropic · Mistral · Ollama.