An AI step is an AI node that does the task, plus the nodes you plug into it: a model (always), and optionally knowledge, memory, tools and an output parser. When an AI step behaves badly, one of these parts is usually the cause.
Reading order
Section titled “Reading order”- Models and where they run: what you plug into an AI node, and cloud vs your computer vs your browser.
- Prompting and outputs: instructions that work, and answers later steps can read.
- Agents and tools: when an agent is worth it, and which one to pick.
- Tool selection and Action approval: what an agent may do, and why it asks you first.
- Knowledge (RAG): answers grounded in your documents, then Embeddings and vectors for how the search works.
- Memory and context, Branching on AI results and Testing AI workflows when you go further.
Three questions to settle first
Section titled “Three questions to settle first”- Which model, where? Pick a cloud model, Ollama on your computer, or a model in your browser. Example Private data stays on-device with Web LLM or Ollama.
- Chain or agent? A chain calls the model once. An agent chooses tools until it's done. Example Summarise a page (chain) vs research a company (agent).
- Its own knowledge or yours? Answers that must come from your documents need a knowledge base. Example A help-desk reply that cites the policy page.
How the parts fit together
Section titled “How the parts fit together”flowchart LR Data[Data from earlier steps] --> Node[AI node: chain or agent] Model[Model] --> Node Knowledge[Local Knowledge] -.-> Node Memory[Chat Memory] -.-> Node Tools[Tool nodes] -.-> Node Parser[Output parser] -.-> Node Node --> Next[Next step] class Node,Model awf-ai class Knowledge,Memory,Parser awf-data class Tools awf-io
Dotted lines are optional. Each part is a node you connect to one of the AI node’s slots on the canvas.
| Part | What it does | Nodes |
|---|---|---|
| AI node | Does the task | Basic LLM Chain, Summarization, Text Classifier, Information Extractor, RAG Agent, Tools Agent |
| Model | Reads and writes text | Chat OpenAI, Chat Anthropic, Ollama, Web LLM and others |
| Knowledge | Your documents, searched by meaning | Local Knowledge, Indexer |
| Memory | The conversation so far | Chat Memory |
| Tools | Actions an agent can choose | Web Search, Browser, HTTP Request, Workflow Tool |
| Output parser | Turns the answer into data | Structured Output Parser, Item List Output Parser |
When AI belongs in a workflow
Section titled “When AI belongs in a workflow”Use AI when the step needs judgment over language or messy content: summarizing a page, sorting messages into categories, pulling fields out of free text. Don’t use AI for work a regular node does more clearly, such as formatting a date, filtering on a known field, or comparing a number with a threshold.
flowchart TD
Need["What the step must do"] --> Known{"Can a fixed rule do it?"}
Known -->|Yes| Deterministic["Use a flow, mapping, code or data node"]
Known -->|No| Language{"Is the input text or images?"}
Language -->|Yes| AI["Use an AI node with focused input"]
Language -->|No| Redesign["Clean up the data first"]
AI --> Check["Check the result before acting on it"]
class Known,Language awf-flow
class Deterministic awf-data
class AI awf-ai
class Check awf-ok
A good AI step usually has:
- one clear task: summarize, classify, extract, answer, or decide;
- clean input from earlier steps (not a whole web page when one section is enough);
- a model that fits the task (see Models and dependencies);
- a structured output when later steps use the result (see Prompting and outputs);
- a plan for empty input, model errors and unexpected answers.
Pick a starting node
Section titled “Pick a starting node”| You want to… | Start with |
|---|---|
| Turn text into other text (rewrite, draft, translate) | Basic LLM Chain |
| Shorten a long text | Summarization |
| Sort text into your own categories | Text Classifier |
| Pull named fields out of text | Information Extractor |
| Check text against a policy before continuing | Guardrails |
| Answer from your own documents | RAG Agent with a knowledge base |
| Let the model choose actions (search, call a service, act on the page) | Tools Agent |
| Keep everything on your device | Local AI nodes |
Common starting points
Section titled “Common starting points”| Goal | Start with | Add next |
|---|---|---|
| Summarize page text | Get All Text → Summarization | A shorter input, a Max words limit |
| Answer questions from your documents | RAG Agent + Local Knowledge | A “not found” answer, testing |
| Research across websites | Tools Agent + Web Search | A small set of tools, a Max iterations limit |
| Let an agent act on a page | Tools Agent + Browser | Action approval |
| Run a chatbot that remembers | Chat Trigger + agent + Chat Memory | Memory and context |
| Produce rows or JSON | Information Extractor + Structured Output Parser | Auto-fixing Output Parser |
Rule of thumb
Section titled “Rule of thumb”If the answer must come from your documents, use a knowledge base. If a later step uses the answer, give it a structure. If the model must choose between actions, use an agent. Before you rely on a workflow, try it with realistic inputs.
Start with Models and where they runWhat you plug into an AI node, and where the model runs.
Practise itRecipes with AI steps, on-device and in the cloud.