Models and dependencies

In 20 seconds

An AI node does the task; the nodes you plug into its slots give it what it needs. Every AI node needs a model. Add memory to remember a conversation, tools to let an agent act, knowledge to answer from your documents, and an output parser to get fields instead of free text. The model can run in the cloud, on your computer, or in your browser.

See it happen

The same Basic LLM Chain reads a signature from a page. Without an output parser, response is text. With a Structured Output Parser plugged in, response holds fields the next step can read, such as {{ $input.response.email }}.

Example run: Get Selected Text, 1 item → Basic LLM Chain, 1 item → Basic LLM Chain + parser, 1 item. Illustration of a run. Field names come from the parser's JSON Output Example.

1. Get Selected Text 1 item Output field name: selection →
2. Basic LLM Chain 1 item no output parser →
3. Basic LLM Chain + parser 1 item Structured Output Parser plugged in

Output of Basic LLM Chain + parser Table JSON

#	response
0	{"name":"Ada Lovelace","email":"ada@example.com","company":"Analytical Engines"}

Illustration of a run. Field names come from the parser's JSON Output Example.

What plugs into an AI node

flowchart LR
  Model[Model] --> Node[AI node]
  Memory[Chat Memory] -.-> Node
  Tools[Tools] -.-> Node
  KB[Local Knowledge] -.-> Node
  Parser[Output parser] -.-> Node
  Input[Data from earlier steps] --> Node --> Next[Next step]
  class Node awf-ai
  class Model awf-ai
  class KB,Memory awf-data
  class Parser awf-data
  class Tools awf-io

Dotted lines are optional. Which slots a node has depends on the node: a chain has no tools slot, for example. See AI agents.

Dependency	What it provides	Example nodes
Chat model	Reads and writes text	Cloud: Chat OpenAI, Chat Anthropic, Chat Google, Chat Mistral, Chat Groq, Chat DeepSeek, Chat xAI, Chat OpenRouter, Chat Azure OpenAI. On your computer: Ollama. In your browser: Web LLM, Transformers Chat, Chrome AI
Knowledge	Searches your documents	Local Knowledge
Memory	Keeps the conversation	Chat Memory
Output parser	Turns the reply into data	Structured Output Parser, Item List Output Parser, Auto-fixing Output Parser
Tool	An action an agent can choose	Web Search, Wikipedia, Browser and others
Embeddings	Turns text into vectors	Local Embeddings, OpenAI Embeddings, Ollama Embeddings and others
Text splitter	Cuts long text into chunks	Character Text Splitter, Recursive Character Text Splitter

Where the model runs

• Cloud → You need the strongest reasoning and reliable tool use. Example Chat OpenAI, Chat Anthropic, Chat Google with your own API key. key in Settings › Providers
• Your computer → Data must stay on your machine, and you can run a model server. Example Ollama running on this computer. private, no per-call cost
• In your browser → Nothing to install, nothing leaves the browser. Example Web LLM, Transformers Chat, Chrome AI. free, smaller, slower

Choose by the job, not by raw power. Try the smallest model that gets your real inputs right, and keep one model per AI step unless you’re comparing them.

You’ll notice this when…

The AI node fails with “Dependency … is required”

A required slot is empty, usually the model. Connect a model node to the AI node’s model slot.

The next step can’t read response.email

Without an output parser, response is plain text. Plug a Structured Output Parser into the AI node and give it a JSON example with the fields you need.

An in-browser model is slow or gives weak answers

In-browser models are small and run on your device. Shorten the input, keep the task narrow, or switch that step to a cloud model. See Local AI.

The Tools Agent doesn’t call any tool

The Tools Agent needs a chat model that supports tool calling. Pick a model that does, and check the tools are connected to its tools slot.

Next

Next concept: Prompting and outputsInstructions that work, and outputs later steps can use.

Practise itRecipes with on-device and cloud AI steps.