openai-assistant-swarm

A simple example

An full example delegating a single input between 3 available assistants...

import OpenAI from 'openai';
import { EnableSwarmAbilities } from '@mintplex-labs/openai-assistant-swarm';
const OpenAIClient = new OpenAI({apiKey: process.env.OPEN_AI_KEY});
EnableSwarmAbilities(OpenAIClient);
await OpenAIClient.beta.assistants.swarm.init();

// Optional - set up listeners here to wait for specific events to return to the user since streaming is not available yet.

// Run the main process on a single text prompt to have work delegate between all of your assistants that are available.
const response = OpenAIClient.beta.assistants.swarm.delegateWithPrompt('What is the weather in New York city right now? Also what is the top stock for today?');
// For example. Given a Pirate bot, Weather Bot, and Stock Bot in your assistant registry on OpenAI.
// Run the below threads in parallel and return to you!
// |--> Will delegate to an existing Weather Bot
// |--> Will delegate to an existing Stock watcher Bot
// -> Pirate bot will not be invoked.
// -----
// The parent will respond with something like "I've arranged for two of our assistants to handle your requests. For assistance with stocks I have delegated that task  to the Stock Bot, and for the weather update in San Francisco, our Weatherbot will provide the current conditions. They will take care of your needs shortly."
//
// You will then get a response once each child responds with either a completion or a `required_action` run you can handle in your codebase easily.

console.log({
  parentRun: response.parentRun, // All information about the parent thread
  subRuns: response.subRuns, // array of runs created and their status for each spun-out child thread!
})

Available tools

Delegation via prompt

First, the main one you are probably interested in - delegation to sub-assistants. Its easy to set up and also to listen to events and add into your current workflow.

// Set up an event listener for when the parent response is completed so you don't have to wait
// for parent + children responses to all complete.
// Useful to return the parent response early while you work on the subtask tool_calls that 
// may or not be required depending on what happened.
OpenAIClient.beta.assistants.swarm.emitter.on('parent_assistant_complete', (args) => {
    console.group('Parent assistant response completed');
    console.log(args.parentRun.playground) // => https://platform.openai.com/playground.... to debug thread & run in browser.
    console.log(args.parentRun.textResponse) // => Yarrh! Want to be speaking to the captain do ya? Ill go fetch them ya land lubber.
    // args.parentRun => The full Run object from OpenAI so you can get the thread_id and other properties like status.
    console.log('\n\n')
    console.groupEnd();
});

// Set up an event listener for when the delegated assistant responses are completed so you don't have to wait
// for parent + children responses to all complete.
// From here you can handle all sub-run tool_calls if they are required to be run.
OpenAIClient.beta.assistants.swarm.emitter.on('child_assistants_complete', (args) => {
    console.group('Child assistant response completed');
    console.log(args.subRuns.map((run) => run.textResponse)) // => Yarrh! I am the captain of this vessel. Ye be after my treasure, Yar?
    console.log(args.subRuns.map((run) => run.playground)) // => https://platform.openai.com/playground.... to debug thread & run in browser.
    // args.subRuns[x].run => The full Run object from OpenAI so you can get the thread_id and other properties like status.
    console.log('\n\n')
    console.groupEnd();
});

// Set up and event listener to see every step event as it is completed:
OpenAIClient.beta.assistants.swarm.emitter.on('poll_event', ({ data }) => {
    console.group('Poll event!');
    console.log({
        status: data.status,
        text: data.prompt || data.textResponse,
        runId: data?.run?.id,
        link: data.playground,
        runStatus: data?.run?.status,
    })
    console.log('\n\n')
    console.groupEnd();
});

// Run the main process on a single text prompt to have work delegate between all of the possible assistants that are available.
const response = OpenAIClient.beta.assistants.swarm.delegateWithPrompt('Let me speak to the head pirate of this vessel! What say ye??');
// You can also just wait for the entire flow to finish instead of setting up listeners to keep the code more synchronous
console.log({
  parentRun: response.parentRun,
  subRuns: response.subRuns,
})

// You can also focus the given task or prompt on a subset of assistants that you know you want to handle delegated work.
// OpenAIClient.beta.assistants.swarm.delegateWithPrompt('Let me speak to the head pirate of this vessel! What say ye??', ['asst_lead_pirate']);

Get all available assistants

Right now, you need to paginate assitants to see who is around to answer a question or handle a task.