Logic & Data Flow

Chains use a powerful templating system to pass information between steps and make complex decisions.

1. Variables & Placeholders

In the configuration of a step, you can access data using the ${...} syntax. The engine supports recursive dot notation to address deeply nested JSON structures.

Accessing Steps

Each step stores its results under its id:

• ${steps.<id>.output}: The raw text response (e.g., from an LLM).
• ${steps.<id>.data.<path>}: Access to the parsed JSON object (e.g., ${steps.weather.data.current.temp}).
• ${steps.<id>.result}: The complete technical result (e.g., an MCP result object).

Global Variables

• ${input} / ${userInput}: The user’s original text.
• ${user_id}: The ID of the current user.
• ${chat_id}: The ID of the current chat.
• ${agent_id}: The ID of the executing agent.

2. Conditions (when)

Use the when field to control whether a step is executed or skipped. The engine evaluates the expression after variable replacement.

Supported Operations:

• Comparisons: == and != (e.g., "${steps.check.data.status} == 'ok'").
• Booleans: Recognizes strings like "true", "false".
• Existence: A non-empty string or a number other than 0 is considered true.

3. Data Flow via Edges (Graphs)

In addition to variable injection, data can also be explicitly mapped via edges. This is useful for injecting specific data into the inputs of a target step:

"edges": [
  {
    "from": "source_step",
    "to": "target_step",
    "map": {
      "data.user.name": "target_field"
    }
  }
]

4. Complex Control Flows

The engine supports advanced structures for logic graphs:

• Branch: Clean conditional branching (switch-case).
• Parallel: Simultaneous execution of branches for performance optimization.
• Loop: Repetition of blocks (iterations).
• Retry: Automatic fault tolerance with adjustable backoff.