Variables and Data Flow Design Techniques

The nodes are connected, but the variables don't flow — this is where beginners most often get stuck. This guide explains how data moves and transforms through a flow.

Sketch the Data Flow First, Then Place the Nodes

Before building, answer three questions:

1. What comes in? Which fields does the Start node collect (user message, custom variables)
2. What goes out? Which data does the Answer node's reply template / the End node's output variables need
3. How does it transform in between? From entry to exit, what processing does the data go through

Write this chain out as a single line, then place nodes against it:

User message → Classification result → (Market branch) Ticker → Market data JSON → Formatted text → Reply

Structured Output: Making LLM Results Usable Downstream

By default the LLM outputs free-form text text (String), from which downstream nodes can hardly extract values precisely. When downstream nodes need to make decisions or pick out fields, turn on the LLM node's structured output switch:

• The output variable becomes structured_output (Object), generated according to the JSON Schema you configure
• When configuring, define each field one by one: parameter name, parameter type, whether it is required
• Typical usage: have the LLM extract {"ticker": "NVDA", "intent": "get quote"}, so downstream IF Else / Http Request nodes can reference the fields directly

Principle: use text for humans, use structured_output for machines.

Division of Labor Among the Three Variable Nodes

Node	What it does	When to use
Variables Transformer	Formatting, type conversion, merging/splitting with Jinja2 templates	Concatenating copy, converting an Object into display text, simple transformations
Parameter Extractor	Uses an LLM to extract structured parameters from text	When natural language needs to be converted into tool-call input parameters
Code	Custom Python / JavaScript logic	Complex computation, cleaning, or validation that a Transformer cannot express

Order of preference: use a Transformer instead of Code whenever possible (templates are easier to maintain than code); Code is the fallback.

Merging Branch Data: Branch Aggregator

When the multiple branches split off by IF Else / Question Classifier each produce different output variables, and a downstream node (such as Answer) only wants to reference "one" result, use the Branch Aggregator:

• Ungrouped mode: aggregates variables from multiple branches into a single output; the type is determined by the first variable selected — so the output types of all branches must be consistent (all String or all Object)
• Grouped mode: use when you need to aggregate multiple groups of variables with different meanings

Anti-pattern: attaching a separate Answer node to each branch and duplicating the same reply template. Instead, use branches → aggregator → a single Answer, so the template is maintained in one place.

Passing Data in Loops

• Iteration: takes a list as input, runs the internal sub-flow once for each element, and produces a list of results. Supports parallel mode for speedup and configurable error handling (a single failed element does not affect the whole). Suitable for batch tasks like "generate a summary for every news article"
• Loop: executes repeatedly by count or condition, updating variables each round and passing them to the next; a special "End Loop" node inside controls the exit. Suitable for tasks like "keep refining until a condition is met"
• Neither can contain Start / End nodes inside; the sub-flow begins from the internal Home node

Troubleshooting Common Data Flow Issues

Symptom	Common cause	Fix
Downstream node can't select the desired variable	Upstream node not connected, or the variable is inside a branch	Check the connections; route branch variables through a Branch Aggregator first
LLM prompt reports "Please fill in the context variable in the prompt"	Context is configured but the variable is not inserted into the prompt	Insert the variable into the prompt, or remove the context configuration
Output JSON fails to parse downstream	Format enforced via prompt only; the model occasionally drifts	Switch to structured output (strict JSON Schema constraints)
Wrong type after aggregation	Branch Aggregator branches have inconsistent types	Unify output types across branches; convert with a Transformer first if needed
Results inside the loop are not carried out	Variables in the loop body are not written back to the iteration output	Check the Iteration output variable configuration