Custom Tool

Watch how to use custom tools

Problem

Function usually takes in structured input data. Let's say you want the LLM to be able to call Airtable Create Record API, the body parameters has to be structured in a specific way. For example:

"records": [
  {
    "fields": {
      "Address": "some address",
      "Name": "some name",
      "Visited": true
    }
  }
]

Ideally, we want LLM to return a proper structured data like this:

{
  "Address": "some address",
  "Name": "some name",
  "Visited": true
}

So we can extract the value and parse it into the body needed for API. However, instructing LLM to output the exact pattern is difficult.

With the new OpenAI Function Calling models, it is now possible. gpt-4-0613 and gpt-3.5-turbo-0613 are specifically trained to return structured data. The model will intelligently choose to output a JSON object containing arguments to call those functions.

Tutorial

Goal: Have the agent automatically get the stock price movement, retrieve related stock news, and add a new record to Airtable.

Let's get started🚀

Create Tools

We need 3 tools to achieve the goal:

Get Stock Price Movement
Get Stock News
Add Airtable Record

Get Stock Price Movement

Create a new Tool with the following details (you can change as you want):

Name: get_stock_movers
Description: Get the stocks that has biggest price/volume moves, e.g. actives, gainers, losers, etc.

Description is an important piece as ChatGPT is relying on this to decide when to use this tool.

JavaScript Function: We are going to use Morning Star /market/v2/get-movers API to get data. First you have to click Subscribe to Test if you haven't already, then copy the code and paste it into JavaScript Function.
- Add const fetch = require('node-fetch'); at the top to import the library. You can import any built-in NodeJS modules and external libraries.
- Return the result at the end.

The final code should be:

const fetch = require('node-fetch');
const url = 'https://morning-star.p.rapidapi.com/market/v2/get-movers';
const options = {
	method: 'GET',
	headers: {
		'X-RapidAPI-Key': 'replace with your api key',
		'X-RapidAPI-Host': 'morning-star.p.rapidapi.com'
	}
};

try {
	const response = await fetch(url, options);
	const result = await response.text();
	console.log(result);
	return result;
} catch (error) {
	console.error(error);
	return '';
}

You can now save it.

Get Stock news

Create a new Tool with the following details (you can change as you want):

Name: get_stock_news
Description: Get latest news for a stock
Output Schema:
- Property: performanceId
- Type: string
- Description: id of the stock, which is referred as performanceID in the API
- Required: true

Output Schema tells ChatGPT what to return as a JSON object. In this case, we are expecting a JSON object like below:

{ "performanceId": "SOME TICKER" }

JavaScript Function: We are going to use Morning Star /news/list API to get the data. First you have to click Subscribe to Test if you haven't already, then copy the code and paste it into JavaScript Function.
- Add const fetch = require('node-fetch'); at the top to import the library. You can import any built-in NodeJS modules and external libraries.
- Return the result at the end.
Next, replace the hard-coded url query parameter performanceId: 0P0000OQN8 to the property variable specified in Output Schema: $performanceId
You can use any properties specified in Output Schema as variables in the JavaScript Function by appending a prefix $ at the front of the variable name.

Final code:

const fetch = require('node-fetch');
const url = 'https://morning-star.p.rapidapi.com/news/list?performanceId=' + $performanceId;
const options = {
	method: 'GET',
	headers: {
		'X-RapidAPI-Key': 'replace with your api key',
		'X-RapidAPI-Host': 'morning-star.p.rapidapi.com'
	}
};

try {
	const response = await fetch(url, options);
	const result = await response.text();
	console.log(result);
	return result;
} catch (error) {
	console.error(error);
	return '';
}

You can now save it.

Add Airtable Record

Create a new Tool with the following details (you can change as you want):

Name: add_airtable
Description: Add the stock, news summary & price move to Airtable
Output Schema:
- Property: stock
- Type: string
- Description: stock ticker
- Required: true
- Property: move
- Type: string
- Description: price move in %
- Required: true
- Property: news_summary
- Type: string
- Description: news summary of the stock
- Required: true

ChatGPT will returns a JSON object like this:

{ "stock": "SOME TICKER", "move": "20%", "news_summary": "Some summary" }

JavaScript Function: We are going to use Airtable Create Record API to create a new record to an existing table. You can find the tableId and baseId from here. You'll also need to create a personal access token, find how to do it here.

Final code should looks like below. Note how we pass in $stock, $move and $news_summary as variables:

const fetch = require('node-fetch');
const baseId = 'your-base-id';
const tableId = 'your-table-id';
const token = 'your-token';

const body = {
	"records": [
		{
			"fields": {
				"stock": $stock,
				"move": $move,
				"news_summary": $news_summary,
			}
		}
	]
};

const options = {
	method: 'POST',
	headers: {
		'Authorization': `Bearer ${token}`,
		'Content-Type': 'application/json'
	},
	body: JSON.stringify(body)
};

const url = `https://api.airtable.com/v0/${baseId}/${tableId}`

try {
	const response = await fetch(url, options);
	const text = await response.text();
	return text;
} catch (error) {
	console.error(error);
	return '';
}

You can now save it.

You should see 3 tools created:

Create Chatflow

You can use the template OpenAI Function Agent from marketplace, and replace the tools with Custom Tool. Select the tool you have created.

Note: OpenAI Function Agent only supports 0613 models currently.

Save the chatflow and start testing it. For starter, you can try asking:

What is the stock that has the biggest price movement today?

The stock that has the biggest price movement today is Overstock.com (OSTK) with a price move of 17.47%.

You can then follow up with another question to get the news of that particular stock:

What are the latest news about this stock that might cause the price movement?

Here are the latest news about Overstock.com (OSTK) that might have caused the price movement:

Title: "Overstock's shares soar on successful 21.5 million bid for Bed Bath & Beyond assets" Source: MarketWatch Published Date: June 22, 2023 Summary: Overstock.com's shares experienced a significant increase after successfully winning a bid for Bed Bath & Beyond assets worth 21.5 million.
Title: "Thinking about trading options or stock in Meta Platforms, Overstock.com, Walmart, Home Depot, or United Parcel Service?" Source: PR Newswire Published Date: June 22, 2023 Summary: This news article discusses the potential trading options and stocks, including Overstock.com, that investors may consider.

Please note that these news articles are for informational purposes only and may not be the sole reason for the price movement. It is always recommended to conduct thorough research and analysis before making any investment decisions.

Lastly, you can ask ChatGPT to add a new record to Airtable:

Can you add a record to Airtable with the stock ticker, price move and news summary?

I have added a record to Airtable with the following details:

Stock Ticker: OSTK Price Move: 17.47% News Summary: Overstock.com's shares experienced a significant increase after successfully winning a bid for Bed Bath & Beyond assets worth $21.5 million.

🎉🎉Voila! That's how you can create your own custom tool and use it with the OpenAI Function Agent!

Additional

Pass Session ID to Function

By default, Function in custom tool has access to the following flow configurations:

$flow.sessionId 
$flow.chatId
$flow.chatflowId
$flow.input

Below is an example of sending the sessionId to Discord webhook:

const fetch = require('node-fetch');
const webhookUrl = "https://discord.com/api/webhooks/1124783587267";
const content = $content; // captured from output schema
const sessionId = $flow.sessionId;

const body = {
	"content": `${mycontent} and the sessionid is ${sessionId}`
};

const options = {
	method: 'POST',
	headers: {
		'Content-Type': 'application/json'
	},
	body: JSON.stringify(body)
};

const url = `${webhookUrl}?wait=true`

try {
	const response = await fetch(url, options);
	const text = await response.text();
	return text;
} catch (error) {
	console.error(error);
	return '';
}

Pass variables to Function

In some cases, you would like to pass variables to custom tool function.

For example, you are creating a chatbot that uses a custom tool. The custom tool is executing a HTTP POST call and API key is needed for successful authenticated request. You can pass it as a variable.

By default, Function in custom tool has access to variables:

$vars.<variable-name>

Example of how to pass variables in Flowise using API and Embedded:

async function query(data) {
    const response = await fetch(
        "http://localhost:3000/api/v1/prediction/<chatflow-id>",
        {
            method: "POST",
            headers: {
                "Content-Type": "application/json"
            },
            body: JSON.stringify(data)
        }
    );
    const result = await response.json();
    return result;
}

query({
    "question": "Hey, how are you?",
    "overrideConfig": {
        "vars": {
            "apiKey": "abc"
        }
    }
}).then((response) => {
    console.log(response);
});

<script type="module">
    import Chatbot from 'https://cdn.jsdelivr.net/npm/flowise-embed/dist/web.js';
    Chatbot.init({
        chatflowid: 'chatflow-id',
        apiHost: 'http://localhost:3000',
        chatflowConfig: {
          vars: {
            apiKey: 'def'
          }
        }
    });
</script>

Example of how to receive the variables in custom tool:

const fetch = require('node-fetch');
const webhookUrl = "https://discord.com/api/webhooks/1124783587267";
const content = $content; // captured from output schema
const sessionId = $flow.sessionId;
const apiKey = $vars.apiKey;

const body = {
	"content": `${mycontent} and the sessionid is ${sessionId}`
};

const options = {
	method: 'POST',
	headers: {
		'Content-Type': 'application/json',
		'Authorization': `Bearer ${apiKey}`
	},
	body: JSON.stringify(body)
};

const url = `${webhookUrl}?wait=true`

try {
	const response = await fetch(url, options);
	const text = await response.text();
	return text;
} catch (error) {
	console.error(error);
	return '';
}

Override Custom Tool

Parameters below can be overriden

Parameter	Description
customToolName	tool name
customToolDesc	tool description
customToolSchema	tool schema
customToolFunc	tool function

Parameter

Description

customToolName

tool name

customToolDesc

tool description

customToolSchema

tool schema

customToolFunc

tool function

Example of an API call to override custom tool parameters:

async function query(data) {
    const response = await fetch(
        "http://localhost:3000/api/v1/prediction/<chatflow-id>",
        {
            method: "POST",
            headers: {
                "Content-Type": "application/json"
            },
            body: JSON.stringify(data)
        }
    );
    const result = await response.json();
    return result;
}

query({
    "question": "Hey, how are you?",
    "overrideConfig": {
        "customToolName": "example_tool",
        "customToolSchema": "z.object({title: z.string()})"
    }
}).then((response) => {
    console.log(response);
});

Import External Dependencies

You can import any built-in NodeJS modules and supported external libraries into Function.

To import any non-supported libraries, you can easily add the new npm package to package.json in packages/components folder.

cd Flowise && cd packages && cd components
pnpm add <your-library>
cd .. && cd ..
pnpm install
pnpm build

Then, add the imported libraries to TOOL_FUNCTION_EXTERNAL_DEP environment variable. Refer #builtin-and-external-dependencies for more details.
Start the app

pnpm start

You can then use the newly added library in the JavaScript Function like so:

const axios = require('axios')

Watch how to add additional dependencies and import libraries

PreviousTools NextVector Stores

Last updated 1 month ago