Value Declarations

In AI JSON, inputs to actions can be:

• links to other actions’ outputs (link)
• external variables (var)
• jinja templates (text)
• lambda functions (lambda)
• environment variables (env)

Link: link

A way to reference the output of another action.
Linked actions will always run before the action that references them.

my_action:
  action: some_action
  an_input: "foo"
another_action:
  action: some_action
  an_input:
    link: my_action

If the output of an action is a pydantic model, you can reference its attributes:

my_action:
  action: some_action
  an_input: "foo"
another_action:
  action: some_action
  an_input:
    link: my_action.an_output_attribute

Variable: var

A way to pass external values directly to action inputs.

my_action:
  action: some_action
  an_input:
    var: my_variable

When running the flow in code, you can pass the value of my_variable as a keyword argument:

from aijson import Flow

async def main():
    flow = Flow.from_file("flow.ai.yaml")
    flow = flow.set_vars(my_variable="foo")
    await flow.run()

Text: text

A way to evaluate jinja templates.

my_action:
  action: some_action
  an_input:
    text: "Hello, {{ name }}"

If a template variable ({{ name }} above) corresponds to another action’s output, it is treated as a link. Otherwise, it is treated as a variable.

If an action’s input is a string, you may omit the text declaration; it will be evaluated as a jinja string by default:

my_action:
  action: some_action
  a_string_input: "Hello, {{ name }}"

Note, AI JSON jinja templates are evaluated in a native jinja environment.

Lambda: lambda

A way to evaluate a python expression.

my_action:
  action: some_action
  an_input:
    lambda: "x + 1"

If a variable (x above) corresponds to another action’s output, it is treated as a link. Otherwise, it is treated as a variable.

Environment: env

A way to reference environment variable values.

my_action:
  action: some_action
  an_input:
    env: MY_ENV_VAR