Creating custom plugins for this project involves extending the Tool class from the langchain/tools module.
Note: I will use the word plugin interchangeably with tool, as the latter is specific to LangChain, and we are mainly conforming to the library.
You are essentially creating DynamicTools in LangChain speak. See the LangChainJS docs for more info.
This guide will walk you through the process of creating your own custom plugins, using the StableDiffusionAPI and WolframAlphaAPI tools as examples.
When using the Functions Agent (the default mode for plugins), tools are converted to OpenAI functions; in any case, plugins/tools are invoked conditionally based on the LLM generating a specific format that we parse.
The most common implementation of a plugin is to make an API call based on the natural language input from the AI, but there is virtually no limit in programmatic use case.
Here are the key takeaways for creating your own plugin:
1. Import Required Modules: Import the necessary modules for your plugin, including the Tool class from langchain/tools and any other modules your plugin might need.
2. Define Your Plugin Class: Define a class for your plugin that extends the Tool class. Set the name and description properties in the constructor. If your plugin requires credentials or other variables, set them from the fields parameter or from a method that retrieves them from your process environment. Note: if your plugin requires long, detailed instructions, you can add a description_for_model property and make description more general.
3. Define Helper Methods: Define helper methods within your class to handle specific tasks if needed.
4. Implement the _call Method: Implement the _call method where the main functionality of your plugin is defined. This method is called when the language model decides to use your plugin. It should take an input parameter and return a result. If an error occurs, the function should return a string representing an error, rather than throwing an error. If your plugin requires multiple inputs from the LLM, read the StructuredTools section.
5. Export Your Plugin and Import into handleTools.js: Export your plugin and import it into handleTools.js. Add your plugin to the toolConstructors object in the loadTools function. If your plugin requires more advanced initialization, add it to the customConstructors object.
6. Export YourPlugin into index.js: Export your plugin into index.js under tools. Add your plugin to the module.exports of the index.js, so you also need to declare it as const in this file.
7. Add Your Plugin to manifest.json: Add your plugin to manifest.json. Follow the strict format for each of the fields of the "plugin" object. If your plugin requires authentication, add those details under authConfig as an array. The pluginKey should match the class name of the Tool class you made, and the authField prop must match the process.env variable name.
Remember, the key to creating a custom plugin is to extend the Tool class and implement the _call method. The _call method is where you define what your plugin does. You can also define helper methods and properties in your class to support the functionality of your plugin.
Note: You can find all the files mentioned in this guide in the .\api\app\langchain\tools folder.
Multi-Input Plugins
If you would like to make a plugin that would benefit from multiple inputs from the LLM, instead of a singular input string as we will review, you need to make a LangChain StructuredTool instead. A detailed guide for this is in progress, but for now, you can look at how I've made StructuredTools in this directory: api\app\clients\tools\structured\. This guide is foundational to understanding StructuredTools, and it's recommended you continue reading to better understand LangChain tools first. The blog linked above is also helpful once you've read through this guide.
Start by importing the necessary modules. This will include the Tool class from langchain/tools and any other modules your tool might need. For example:
const { Tool } = require('langchain/tools');
// ... whatever else you needNext, define a class for your plugin that extends the Tool class. The class should have a constructor that calls the super() method and sets the name and description properties. These properties will be used by the language model to determine when to call your tool and with what parameters.
Important: you should set credentials/necessary variables from the fields parameter, or alternatively from a method that gets it from your process environment
class StableDiffusionAPI extends Tool {
constructor(fields) {
super();
this.name = 'stable-diffusion';
this.url = fields.SD_WEBUI_URL || this.getServerURL(); // <--- important!
this.description = `You can generate images with 'stable-diffusion'. This tool is exclusively for visual content...`;
}
...
}Optional: As of v0.5.8, when using Functions, you can add longer, more detailed instructions, with the description_for_model property. When doing so, it's recommended you make the description property more generalized to optimize tokens. Each line in this property is prefixed with // to mirror how the prompt is generated for ChatGPT (chat.openai.com). This format more closely aligns to the prompt engineering of official ChatGPT plugins.
// ...
this.description_for_model = `// Generate images and visuals using text with 'stable-diffusion'.
// Guidelines:
// - ALWAYS use {{"prompt": "7+ detailed keywords", "negative_prompt": "7+ detailed keywords"}} structure for queries.
// - Visually describe the moods, details, structures, styles, and/or proportions of the image. Remember, the focus is on visual attributes.
// - Craft your input by "showing" and not "telling" the imagery. Think in terms of what you'd want to see in a photograph or a painting.
// - Here's an example for generating a realistic portrait photo of a man:
// "prompt":"photo of a man in black clothes, half body, high detailed skin, coastline, overcast weather, wind, waves, 8k uhd, dslr, soft lighting, high quality, film grain, Fujifilm XT3"
// "negative_prompt":"semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime, out of frame, low quality, ugly, mutation, deformed"
// - Generate images only once per human query unless explicitly requested by the user`;
this.description = 'You can generate images using text with \'stable-diffusion\'. This tool is exclusively for visual content.';
// ...Within the constructor, note that we're getting a sensitive variable from either the fields object or from the getServerURL method we define to access an environment variable.
this.url = fields.SD_WEBUI_URL || this.getServerURL();Any credentials necessary are passed through fields when the user provides it from the frontend; otherwise, the admin can "authorize" the plugin for all users through environment variables. All credentials passed from the frontend are encrypted.
// It's recommended you follow this convention when accessing environment variables.
getServerURL() {
const url = process.env.SD_WEBUI_URL || '';
if (!url) {
throw new Error('Missing SD_WEBUI_URL environment variable.');
}
return url;
}You can define helper methods within your class to handle specific tasks if needed. For example, the StableDiffusionAPI class includes methods like replaceNewLinesWithSpaces, getMarkdownImageUrl, and getServerURL to handle various tasks.
class StableDiffusionAPI extends Tool {
...
replaceNewLinesWithSpaces(inputString) {
return inputString.replace(/\r\n|\r|\n/g, ' ');
}
...
}The _call method is where the main functionality of your plugin is implemented. This method is called when the language model decides to use your plugin. It should take an input parameter and return a result.
In a basic Tool, the LLM will generate one string value as an input. If your plugin requires multiple inputs from the LLM, read the StructuredTools section.
class StableDiffusionAPI extends Tool {
...
async _call(input) {
// Your tool's functionality goes here
...
return this.result;
}
}Important: The _call function is what will the agent will actually call. When an error occurs, the function should, when possible, return a string representing an error, rather than throwing an error. This allows the error to be passed to the LLM and the LLM can decide how to handle it. If an error is thrown, then execution of the agent will stop.
This process will be somewhat automated in the future, as long as you have your plugin/tool in api\app\langchain\tools
// Export
module.exports = StableDiffusionAPI;/* api\app\langchain\tools\handleTools.js */
const StableDiffusionAPI = require('./StableDiffusion');
...In handleTools.js, find the beginning of the loadTools function and add your plugin/tool to the toolConstructors object.
const loadTools = async ({ user, model, tools = [], options = {} }) => {
const toolConstructors = {
calculator: Calculator,
google: GoogleSearchAPI,
wolfram: WolframAlphaAPI,
'dall-e': OpenAICreateImage,
'stable-diffusion': StableDiffusionAPI // <----- Newly Added. Note: the key is the 'name' provided in the class.
// We will now refer to this name as the `pluginKey`
};If your Tool class requires more advanced initialization, you would add it to the customConstructors object.
The default initialization can be seen in the loadToolWithAuth function, and most custom plugins should be initialized this way.
Here are a few customConstructors, which have varying initializations
const customConstructors = {
browser: async () => {
let openAIApiKey = process.env.OPENAI_API_KEY;
if (!openAIApiKey) {
openAIApiKey = await getUserPluginAuthValue(user, 'OPENAI_API_KEY');
}
return new WebBrowser({ model, embeddings: new OpenAIEmbeddings({ openAIApiKey }) });
},
// ...
plugins: async () => {
return [
new HttpRequestTool(),
await AIPluginTool.fromPluginUrl(
"https://www.klarna.com/.well-known/ai-plugin.json", new ChatOpenAI({ openAIApiKey: options.openAIApiKey, temperature: 0 })
),
]
}
};Find the index.js under api/app/clients/tools. You need to put your plugin into the module.exports, to make it compile, you will also need to declare your plugin as consts:
const StructuredSD = require('./structured/StableDiffusion');
const StableDiffusionAPI = require('./StableDiffusion');
...
module.exports = {
...
StableDiffusionAPI,
StructuredSD,
...
}This process will be somehwat automated in the future along with step 5, as long as you have your plugin/tool in api\app\langchain\tools, and your plugin can be initialized with the default method
{
"name": "Calculator",
"pluginKey": "calculator",
"description": "Perform simple and complex mathematical calculations.",
"icon": "https://i.imgur.com/RHsSG5h.png",
"isAuthRequired": "false",
"authConfig": []
},
{
"name": "Stable Diffusion",
"pluginKey": "stable-diffusion",
"description": "Generate photo-realistic images given any text input.",
"icon": "https://i.imgur.com/Yr466dp.png",
"authConfig": [
{
"authField": "SD_WEBUI_URL",
"label": "Your Stable Diffusion WebUI API URL",
"description": "You need to provide the URL of your Stable Diffusion WebUI API. For instructions on how to obtain this, see <a href='url'>Our Docs</a>."
}
]
},Each of the fields of the "plugin" object are important. Follow this format strictly. If your plugin requires authentication, you will add those details under authConfig as an array since there could be multiple authentication variables. See the Calculator plugin for an example of one that doesn't require authentication, where the authConfig is an empty array (an array is always required).
Note: as mentioned earlier, the pluginKey matches the class name of the Tool class you made.
Note: the authField prop must match the process.env variable name
Here is an example of a plugin with more than one credential variable
[
{
"name": "Google",
"pluginKey": "google",
"description": "Use Google Search to find information about the weather, news, sports, and more.",
"icon": "https://i.imgur.com/SMmVkNB.png",
"authConfig": [
{
"authField": "GOOGLE_CSE_ID",
"label": "Google CSE ID",
"description": "This is your Google Custom Search Engine ID. For instructions on how to obtain this, see <a href='https://github.com/danny-avila/LibreChat/blob/main/docs/features/plugins/google_search.md'>Our Docs</a>."
},
{
"authField": "GOOGLE_API_KEY",
"label": "Google API Key",
"description": "This is your Google Custom Search API Key. For instructions on how to obtain this, see <a href='https://github.com/danny-avila/LibreChat/blob/main/docs/features/plugins/google_search.md'>Our Docs</a>."
}
]
},Here's another example of a custom tool, the WolframAlphaAPI tool. This tool uses the axios module to make HTTP requests to the Wolfram Alpha API.
const axios = require('axios');
const { Tool } = require('langchain/tools');
class WolframAlphaAPI extends Tool {
constructor(fields) {
super();
this.name = 'wolfram';
this.apiKey = fields.WOLFRAM_APP_ID || this.getAppId();
this.description = `Access computation, math, curated knowledge & real-time data through wolframAlpha...`;
}
async fetchRawText(url) {
try {
const response = await axios.get(url, { responseType: 'text' });
return response.data;
} catch (error) {
console.error(`Error fetching raw text: ${error}`);
throw error
}
}
getAppId() {
const appId = process.env.WOLFRAM_APP_ID || '';
if (!appId) {
throw new Error('Missing WOLFRAM_APP_ID environment variable.');
}
return appId;
}
createWolframAlphaURL(query) {
const formattedQuery = query.replaceAll(/`/g, '').replaceAll(/\n/g, ' ');
const baseURL = 'https://www.wolframalpha.com/api/v1/llm-api';
const encodedQuery = encodeURIComponent(formattedQuery);
const appId = this.apiKey || this.getAppId();
const url = `${baseURL}?input=${encodedQuery}&appid=${appId}`;
return url;
}
async _call(input) {
try {
const url = this.createWolframAlphaURL(input);
const response = await this.fetchRawText(url);
return response;
} catch (error) {
if (error.response && error.response.data) {
console.log('Error data:', error.response.data);
return error.response.data;
} else {
console.log(`Error querying Wolfram Alpha`, error.message);
return 'There was an error querying Wolfram Alpha.';
}
}
}
}
module.exports = WolframAlphaAPI;In this example, the WolframAlphaAPI class has helper methods like fetchRawText, getAppId, and createWolframAlphaURL to handle specific tasks. The _call method makes an HTTP request to the Wolfram Alpha API and returns the response.