Visual Flows Guide

Flows let you chain multiple AI operations together in a visual graph editor powered by Node-RED. Each node in your flow performs one action — calling an MCP tool, running TypeScript code, prompting an LLM, making a decision, or waiting for human input. Data flows from one node to the next automatically.

OpenShopFloor uses a hybrid execution model: custom OSF nodes run through the optimized OSF engine with SSE streaming, while all native Node-RED nodes (switch, change, function, http-request, etc.) are fully supported.

1. Open the Editor — Go to Flows and click "Open Editor". The Node-RED editor opens in a full-screen view.
2. Add Nodes — Drag nodes from the left palette onto the canvas. Look for the "OpenShopFloor" category for custom OSF nodes.
3. Connect Nodes — Draw wires between node outputs and inputs to define the execution order.
4. Configure — Double-click a node to set its parameters (which tool to call, template text, LLM settings, etc.).
5. Save — Click the Save button in the top bar. Give your flow a name and description.
6. Run — Go back to the Flows listing and click Run on your saved flow. Results stream in real time.

Some nodes support multiple output ports, letting you route data to different downstream paths:

• osf-ts — Configure 1-5 outputs. Return an array where each element goes to the corresponding port.
• osf-decision — Each condition maps to a separate output port.
• switch — Each rule creates an output port for conditional routing.
• function — Return an array of messages for multiple outputs.

Nodes like osf-context and osf-llm accept multiple inputs. The engine waits for all upstream nodes to complete before executing a multi-input node.

• osf-context — Merges all upstream outputs into a single JSON object, keyed by source node label.
• osf-llm — Expects two inputs: context (system message from osf-context) and prompt (user message from osf-prompt-tpl).
• join — Collects multiple messages and combines them into an array or object.

1. The flow engine builds a directed acyclic graph (DAG) from your nodes and wires.
2. Entry nodes (no incoming connections) execute first.
3. Each node receives the output of its predecessor as context. Multi-input nodes receive all upstream outputs.
4. Execution proceeds in topological order — all dependencies must complete before a node runs.
5. Multi-output nodes route data to specific downstream paths based on port index.
6. If a human-input node is reached, the flow pauses until the user responds.
7. Results are streamed in real time via SSE (Server-Sent Events).

A modular flow that collects data from multiple sources, builds context, and generates an analysis:

A flow that monitors defects and escalates to a human when quality drops:

• Start simple — begin with 2-3 nodes and add complexity gradually.
• Use osf-context for multi-source data — collect data from multiple MCP/TS nodes before sending to the LLM.
• Use osf-prompt-tpl for structured prompts — separate your prompt template from data collection.
• Add human checkpoints — use human-input nodes before destructive or high-impact actions.
• Validate LLM output — use osf-output-parser after osf-llm to ensure structured JSON responses.
• Use debug nodes — attach debug nodes to inspect data at any point in the flow.
• Name your nodes — double-click and set a descriptive name. osf-context uses these names as JSON keys.
• One tab per flow — each Node-RED tab becomes a separate saveable flow.