Tools

Fm::Tool

An abstract class for defining tools that the model can call during generation.

Abstract Methods

Subclasses must implement all four methods:

class MyTool < Fm::Tool
  def name : String
    "myTool"
  end

  def description : String
    "Description of what this tool does"
  end

  def arguments_schema : JSON::Any
    JSON.parse(%({"type":"object","properties":{...},"required":[...]}))
  end

  def call(arguments : JSON::Any) : Fm::ToolOutput
    # Execute and return result
    Fm::ToolOutput.new("result")
  end
end

Method	Return Type	Description
`name`	`String`	Unique identifier for the tool
`description`	`String`	Describes what the tool does
`arguments_schema`	`JSON::Any`	JSON Schema for the tool's input
`call(arguments)`	`ToolOutput`	Executes the tool with parsed arguments

Class Methods

`.tools_to_json`

Fm::Tool.tools_to_json(tools : Array(Fm::Tool)) : String

Serializes an array of tool definitions to a JSON string for the FFI layer. This is used internally when creating sessions with tools.

Fm::ToolOutput

A struct wrapping the result of a tool invocation.

Constructors

# From a string
Fm::ToolOutput.new(content : String)

# From JSON
Fm::ToolOutput.new(json: JSON::Any)

Constructor	Description
`.new(content)`	Creates output from a plain string
`.new(json:)`	Creates output from a `JSON::Any` value (serialized to string)

Properties

Property	Type	Description
`content`	`String`	The tool's output content

Arguments Schema

The arguments_schema should return a valid JSON Schema object describing the tool's expected input. The model uses this schema to construct valid arguments.

def arguments_schema : JSON::Any
  JSON.parse(<<-JSON
    {
      "type": "object",
      "properties": {
        "query": {
          "type": "string",
          "description": "The search query"
        },
        "limit": {
          "type": "integer",
          "description": "Maximum number of results"
        }
      },
      "required": ["query"]
    }
  JSON
  )
end