johannlai/openai-function-calling-tools

🛠 openai function calling tools for JS/TS

OpenAI Function calling tools

OpenAI Function calling tools

OpenAI Function calling tools is a repository that offers a set of tools to help you easy to build a function calling model with OpenAI API.

More information about function calling

Sample: https://chatFn.io

🪓 Tools

The repo provides the following tools you can use out of the box:

• 🗺️ ShowPoisOnMap: A tool that can show points of interest on a map.
• 🌐 ReverseGeocode: A tool that can convert coordinates into a human-readable address.
• ⏰ Clock: A clock that can tell you the time.
• 🧮 Calculator: A simple calculator that can do basic arithmetic. Input should be a math expression.
• 🔍 GoogleCustomSearch: A wrapper around the Google Custom Search API. Useful for when you need to answer questions about current events. Input should be a search query.
• 🔍 BingCustomSearch: A wrapper around the Bing Custom Search API. Useful for when you need to answer questions about current events. Input should be a search query.
• 🔍 SerperCustomSearch: A wrapper around the SerpAPI. Useful for when you need to answer questions about current events. Input should be a search query.
• 🏞️ SerperImagesSearch: Use SerpAPI to search images. Input should be a search query.
• 📁 fs: WriteFileTool abd ReadFileTool access to the file system. Input should be a file path and text written to the file.
• 🪩 webbrowser: A web browser that can open a website. Input should be a URL.
• 🚧 sql: Input to this tool is a detailed and correct SQL query, output is a result from the database.
• 🚧 JavaScriptInterpreter: A JavaScript interpreter. Input should be a JavaScript program string.

You can use { Tool } factory function to create a tool instance. See /tools for more examples.

📦 Quick Install

npm install openai-function-calling-tools

📖 Usage

Example 1: Function Calls

use JavaScriptInterpreter to calculate 0.1 + 0.2

import { Configuration, OpenAIApi } from "openai";
import { createCalculator } from "openai-function-calling-tools"

const configuration = new Configuration({
    apiKey: process.env.OPENAI_API_KEY,
  });
const openai = new OpenAIApi(configuration);

const QUESTION = "What is 100*2?";

const messages = [
  {
    role: "user",
    content: QUESTION,
  },
];

# ✨ STEP 1: new the tools you want to use
const [calculator, calculatorSchema] = createCalculator();

# ✨ STEP 2:  add the tools to the functions object
const functions = {
  calculator,
};

const getCompletion = async (messages) => {
  const response = await openai.createChatCompletion({
    model: "gpt-3.5-turbo-0613",
    messages,
    # ✨ STEP 3: add the tools to the schema
    functions: [calculatorSchema],
    temperature: 0,
  });

  return response;
};

console.log("Question: " + QUESTION);
let response = await getCompletion(messages);

if (response.data.choices[0].finish_reason === "function_call") {
  const fnName = response.data.choices[0].message.function_call.name;
  const args = response.data.choices[0].message.function_call.arguments;

  console.log("Function call: " + fnName);
  console.log("Arguments: " + args);

  # ✨ STEP 4: call the function
  const fn = functions[fnName];
  const result = fn(JSON.parse(args));

  console.log("Calling Function Result: " + result);

  messages.push({
    role: "assistant",
    content: null,
    function_call: {
      name: fnName,
      arguments: args,
    },
  });

  messages.push({
    role: "function",
    name: fnName,
    content: JSON.stringify({ result: result }),
  });

  // call the completion again
  response = await getCompletion(messages);

  console.log(response.data.choices[0].message.content);
}

Example 2: Function Calls with Google Custom Search

📝 Note: You need to apply for a Google Custom Search API key and a Google Custom Search Engine ID to use this tool.

The following is a sequence diagram of the example

sequenceDiagram
  participant U as User
  participant M as Main Function
  participant O as OpenAI API
  participant F as Functions Object
  participant GC as Google Custom Search

  U->>M: Execute main function
  M->>M: Initialize configuration and API
  M->>M: Define QUESTION variable
  M->>M: Create Google Custom Search tool
  M->>F: Add tool to functions object
  loop Chat Completion Loop
      M->>O: Request chat completion
      O-->>M: Return response
      alt If finish reason is "stop"
          M->>U: Display answer and exit loop
      else If finish reason is "function_call"
          M->>M: Parse function call name and arguments
          M->>F: Invoke corresponding function
          F->>GC: Perform Google Custom Search
          GC-->>F: Return search results
          F->>M: Receive function result
          M->>M: Add result to message queue
          M->>M: Output function call details
      else Other cases
          M->>M: Continue loop
      end
  end

Code

const { Configuration, OpenAIApi } = require("openai");
const { createGoogleCustomSearch } = require("openai-function-calling-tools");

const main = async () => {
  const configuration = new Configuration({
    apiKey: process.env.OPENAI_API_KEY,
  });
  const openai = new OpenAIApi(configuration);

  const QUESTION = "How many tesla model 3 sale in 2022?"

  const messages = [
    {
      role: "user",
      content: QUESTION,
    },
  ];

  // ✨ STEP 1: new the tools you want to use
  const [googleCustomSearch, googleCustomSearchSchema] =
    createGoogleCustomSearch({
      apiKey: process.env.GOOGLE_API_KEY,
      googleCSEId: process.env.GOOGLE_CSE_ID,
    });


  // ✨ STEP 2:  add the tools to the functions object
  const functions = {
    googleCustomSearch,
  };

  const getCompletion = async (messages) => {
    const response = await openai.createChatCompletion({
      model: "gpt-3.5-turbo-0613",
      messages,
      // ✨ STEP 3: add the tools schema to the functions parameter
      functions: [googleCustomSearchSchema],
      temperature: 0,
    });

    return response;
  };
  let response;

  console.log("Question: " + QUESTION);

  while (true) {
    response = await getCompletion(messages);

    if (response.data.choices[0].finish_reason === "stop") {
      console.log(response.data.choices[0].message.content);
      break;
    } else if (response.data.choices[0].finish_reason === "function_call") {
      const fnName = response.data.choices[0].message.function_call.name;
      const args = response.data.choices[0].message.function_call.arguments;

      const fn = functions[fnName];
      const result = await fn(JSON.parse(args));

      console.log(`Function call: ${fnName}, Arguments: ${args}`);
      console.log(`Calling Function ${fnName} Result: ` + result);

      messages.push({
        role: "assistant",
        content: "",
        function_call: {
          name: fnName,
          arguments: args,
        },
      });

      messages.push({
        role: "function",
        name: fnName,
        content: JSON.stringify({ result: result }),
      });
    }
  }
};

main();

Example 3: Schema Extraction

Example to extract schema from a function call

Tree structure:

import { Configuration, OpenAIApi } from "openai";

const configuration = new Configuration({
  apiKey: process.env.OPENAI_API_KEY,
});
const openai = new OpenAIApi(configuration);

const getCompletion = async (messages) => {
  const response = await openai.createChatCompletion({
    model: "gpt-3.5-turbo-0613",
    messages: [
      {
        role: "user",
        content: `root
              ├── folder1
              │   ├── file1.txt
              │   └── file2.txt
              └── folder2
                  ├── file3.txt
                      └── subfolder1
                              └── file4.txt`
      },
    ],
    functions: [
      {
        "name": "buildTree",
        "description": "build a tree structure",
        "parameters": {
          "type": "object",
          "properties": {
            "name": {
              "type": "string",
              "description": "The name of the node"
            },
            "children": {
              "type": "array",
              "description": "The tree nodes",
              "items": {
                "$ref": "#"
              }
            },
            "type": {
              "type": "string",
              "description": "The type of the node",
              "enum": [
                "file",
                "folder"
              ]
            }
          },
          "required": [
            "name",
            "children",
            "type"
          ]
        }
      }
    ],
    temperature: 0,
  });

  return response;
};

let response = await getCompletion();

if (response.data.choices[0].finish_reason === "function_call") {
  const args = response.data.choices[0].message.function_call.arguments;
  // 🌟 output the Tree structure data
  console.log(args);
}

💻 Supported Environments

• Node.js v16 or higher
• Cloudflare Workers
• Vercel / Next.js (Backend, Serverless and Edge functions 🔥)
• Supabase Edge Functions
• 🚧 Browser

🛡️ Safe for Production

🌟 Inspiration

• LangChainAI

openai-function-calling-tools's Issues

vm2 is deprecated for critical security issue Please Update

i am getting this error :

npm WARN deprecated [email protected]: The library contains critical security issues and should not be used for production! The maintenance of the project has been discontinued. Consider migrating your code to isolated-vm.

Remove Axios from this lib

Because Axios is reporting an error the edge environment: Error: Adapter 'http' is not available in the build.

refer: axios/axios#5523

README里示例代码import错误

import { googleCustomSearch } from ...
应为
import { GoogleCustomSearch } from ...

import `openai-function-calling-tools` 模块后报错 Module not found.

如图，移除 import 不报错，加上就报错.

How do you pass a context object to a function call?

for example, a function to store contact info:

const saveContact = (name, phoneNumber) => {...}

in the runtime, the call should be:

const saveContact = (sessionCtx, name, phoneNumber) => { ... }

I think OpenAI don't need to know about session context?

Implement a function call based on TypeChat that works with all LLMs?

Similar to Microsoft's TypeChat, is it possible to implement a function call based on this that works with all LLMs?

Anyscale compatibility

Can you make it work with Mixtral or LLaMA from Anyscale?
https://docs.endpoints.anyscale.com/guides/function-calling

Export Tool type and let user create more tools

export Tool and make it async. then we can replace Langchainjs, loooool

Design a pattern for running JavaScriptInterpreter on Cloudflare.

to Docker, Cloudflare Worker can also achieve similar effects, quickly starting and stopping a script project.