@draftor/tools

A simple TypeScript npm package designed to transform your function comments into an OpenAI Tool calling format. It's an alternative to zod and z. Built to maintain functions used for tool calling in a separate file, this package allows you to generate OpenAI Tool Calling JSON format without writing extensive code or using .describe().

⚠️ Caution: Check your Linter and Prettier config before add this package!

This approach deviates from typical TypeScript patterns. The comment must reside within the function, not outside. For example:

DO ✅ 
export function foo(bar:string) {
  /**
   * @description Converts a number to its string representation.
   * @param {boolean} bar - The bool to convert.
   * @param {number} abc - The number to convert.
   * @param {string} xyz - The string to convert.
   * @param {undefined} mpn - The any to convert.
   * @returns {object} The Object as response.
   */
  return bar.toString();
}

DON'T ❌
/**
 * @description Converts a number to its string representation.
 * @param {boolean} bar - The number to convert.
 * @param {number} abc - The number to convert.
 * @param {string} xyz - The number to convert.
 * @param {undefined} mpn - The number to convert.
 * @returns {object} The string representation of the input number.
 */
export function foo(bar: string) {
  return bar.toString();
}

Installation

To install the package, use npm:

npm install @draftor/tools

Usage

1. Converting Functions to OpenAI Format

Here's a basic example of how to use the Tools class:

Your tool/functions for tool calling

# yourFunctions.ts

export function foo(bar:string) {
  /**
   * @description Converts a number to its string representation.
   * @param {boolean} bar - The number to convert.
   * @param {number} abc - The number to convert.
   * @param {string} xyz - The number to convert.
   * @param {undefined} mpn - The number to convert.
   * @returns {object} The string representation of the input number.
   */
  return bar.toString();
}

import { Tools } from '@draftor/tools';
import { foo } from './yourFunctions';

const tools = new Tools(funcs);
const result = tools.toOpenAI(); // --> for object response
// const result = tools.toOpenAI('string'); -->  for string response

console.log(result); //will print in json string as output

{
  "name": "foo",
  "description": "No description provided.",
  "params": {
    "type": "object",
    "properties": {
      "bar": {
        "type": "boolean",
        "description": "The number to convert."
      },
      "abc": {
        "type": "number",
        "description": "The number to convert."
      },
      "xyz": {
        "type": "string",
        "description": "The number to convert."
      },
      "mpn": {
        "type": "any",
        "description": "The number to convert."
      }
    },
    "required": [
      "bar",
      "abc",
      "xyz",
      "mpn"
    ]
  },
  "returns": {
    "type": "object",
    "description": "The string representation of the input number."
  }
}

2. Executing LLM ToolCalls from response

// Response Format

export interface IToolCall {
  index: number,
  id: string;
  type: 'function';
  function: IFunctionCall;
}

export interface IFunctionCall {
  name: string;
  arguments: string;
}

How to execute the tool calls

const response = await LLM.chat({msg, tools}) // mock api. Use an endpoint of your choice

const toolsFromLLM = getToolsFromResponse(response) as IToolCall; // Implement response.data.choices[0].message.content logic with or without stream and extract tools object.

/**
 * You can either use the ToolCall response directly by looping over the tools and executing them.
 * ---===  OR  ===---
 * Implement the logic yourself, convert the tools to functions and arguments, and pass them to this function for execution.
 * However, if you've already implemented this much, it's pretty unnecessary to use the .exec() function!
 */

const funResp = tool.exec(toolsFromLLM); // If there is code in the args response, be sure to have an escape logic, but Ideally shoould work.
 
 OR
const funResp = tool.exec('foo', { bar: 'Waba laba dub dub!' });

Contributing

Contributions are welcome! Please feel free to submit a pull request or open an issue.

License

This project is licensed under the MIT License.

Author

https://x.com/p_naix

Acknowledgments

• doctrine for JSDoc parsing.