Function Call Schema in AI Module

The AI module provides a robust mechanism for defining function call schemas that LLMs can use to interact with your code. This document explains the schema options available for more precise value constraints.

Constraints

When defining function call parameters through context definitions, you can use the following constraints:

Constant Values

Use the FixedValue constraint when a parameter must have a specific value that the LLM cannot change. This is useful for enforcing specific identifiers or settings.

'entity_type' => new ContextDefinition(
  data_type: 'string',
  label: 'Entity Type',
  description: 'The entity type to interact with.',
  required: TRUE,
  constraints: [
    'FixedValue' => 'node',
  ],
),

The constraint will be translated to a const property in the JSON schema, indicating to the LLM that this value is fixed and cannot be changed.

Allowed Values

For parameters that must be one of a predefined set of values, use the Choice constraints.

'bundle' => new ContextDefinition(
  data_type: 'string',
  label: 'Content Type',
  description: 'The content type to use.',
  required: TRUE,
  constraints: [
    'Choice' => ['article', 'page', 'blog'],
  ],
),

These constraints are rendered as an enum in the JSON schema, providing the LLM with a list of valid options.

Length Constraints

For string parameters where you need to enforce minimum and/or maximum length, use the Length constraint.

'title' => new ContextDefinition(
  data_type: 'string',
  label: 'Title',
  description: 'The title of the content item.',
  required: TRUE,
  constraints: [
    'Length' => [
      'min' => 3,
      'max' => 255,
    ],
  ],
),

This ensures the LLM generates strings within the specified length range.

Numeric Range

For numeric parameters where values must fall within specific bounds, use the Range constraint.

'weight' => new ContextDefinition(
  data_type: 'integer',
  label: 'Weight',
  description: 'The weight determining the order of the item (lower values appear first).',
  required: TRUE,
  constraints: [
    'Range' => [
      'min' => -100,
      'max' => 100,
    ],
  ],
),

This ensures the LLM generates numeric values within the specified range.

Example Plugin with Constraints

Here's a complete example of a function call plugin using various constraints:

#[FunctionCall(
  id: 'my_module:create_content',
  function_name: 'create_content',
  name: 'Create Content',
  description: 'Creates a new content item with specified parameters.',
  group: 'content_management',
  context_definitions: [
    'entity_type' => new ContextDefinition(
      data_type: 'string',
      label: 'Entity Type',
      description: 'The entity type (always "node").',
      required: TRUE,
      constraints: [
        'FixedValue' => 'node',
      ],
    ),
    'bundle' => new ContextDefinition(
      data_type: 'string',
      label: 'Content Type',
      description: 'The content type to create.',
      required: TRUE,
      constraints: [
        'Choice' => ['article', 'page', 'blog'],
      ],
    ),
    'title' => new ContextDefinition(
      data_type: 'string',
      label: 'Title',
      description: 'The title of the content.',
      required: TRUE,
      constraints: [
        'Length' => ['min' => 3, 'max' => 255],
      ],
    ),
    'status' => new ContextDefinition(
      data_type: 'boolean',
      label: 'Published',
      description: 'Whether the content should be published.',
      required: FALSE,
      default_value: TRUE,
    ),
  ],
)]

By using these constraints, you ensure that LLMs provide more accurate and valid parameters to your function calls, reducing errors and improving reliability.