OpenAI AssistantsStream Assistant Responses with Tools

Stream Assistant Responses with Tools

Let's create a simple chat interface that allows users to send messages to the assistant and receive responses and give it the ability to use tools. You will integrate the useAssistant hook from ai/react to stream the messages and status.

You will need to provide the list of tools on the OpenAI Assistant Dashboard. You can use the following schema to create a tool to convert celsius to fahrenheit.

{
  "name": "celsiusToFahrenheit",
  "description": "convert celsius to fahrenheit.",
  "parameters": {
    "type": "object",
    "properties": {
      "value": {
        "type": "number",
        "description": "the value in celsius."
      }
    },
    "required": ["value"]
  }
}

Client

Let's create a simple chat interface that allows users to send messages to the assistant and receive responses. You will integrate the useAssistant hook from ai/react to stream the messages and status.

pages/index.tsx

import { Message, useAssistant } from 'ai/react';

export default function Page() {
  const { status, messages, input, submitMessage, handleInputChange } =
    useAssistant({ api: '/api/assistant' });

  return (
    <div className="flex flex-col gap-2">
      <div className="p-2">status: {status}</div>

      <div className="flex flex-col p-2 gap-2">
        {messages.map((message: Message) => (
          <div key={message.id} className="flex flex-row gap-2">
            <div className="w-24 text-zinc-500">{`${message.role}: `}</div>
            <div className="w-full">{message.content}</div>
          </div>
        ))}
      </div>

      <form onSubmit={submitMessage} className="fixed bottom-0 p-2 w-full">
        <input
          disabled={status !== 'awaiting_message'}
          value={input}
          onChange={handleInputChange}
          className="bg-zinc-100 w-full p-2"
        />
      </form>
    </div>
  );
}

Server

Next, you will create an API route for api/assistant to handle the assistant's messages and responses. You will use the AssistantResponse function from ai to stream the assistant's responses back to the useAssistant hook on the client.

app/api/assistant/route.ts

import { AssistantResponse } from 'ai';
import OpenAI from 'openai';

const openai = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY || '',
});

export async function POST(req: Request) {
  const input: {
    threadId: string | null;
    message: string;
  } = await req.json();

  const threadId = input.threadId ?? (await openai.beta.threads.create({})).id;

  const createdMessage = await openai.beta.threads.messages.create(threadId, {
    role: 'user',
    content: input.message,
  });

  return AssistantResponse(
    { threadId, messageId: createdMessage.id },
    async ({ forwardStream }) => {
      const runStream = openai.beta.threads.runs.stream(threadId, {
        assistant_id:
          process.env.ASSISTANT_ID ??
          (() => {
            throw new Error('ASSISTANT_ID is not set');
          })(),
      });

      let runResult = await forwardStream(runStream);

      while (
        runResult?.status === 'requires_action' &&
        runResult.required_action?.type === 'submit_tool_outputs'
      ) {
        const tool_outputs =
          runResult.required_action.submit_tool_outputs.tool_calls.map(
            (toolCall: any) => {
              const parameters = JSON.parse(toolCall.function.arguments);

              switch (toolCall.function.name) {
                case 'celsiusToFahrenheit':
                  const celsius = parseFloat(parameters.value);
                  const fahrenheit = celsius * (9 / 5) + 32;

                  return {
                    tool_call_id: toolCall.id,
                    output: `${celsius}°C is ${fahrenheit.toFixed(2)}°F`,
                  };

                default:
                  throw new Error(
                    `Unknown tool call function: ${toolCall.function.name}`,
                  );
              }
            },
          );

        runResult = await forwardStream(
          openai.beta.threads.runs.submitToolOutputsStream(
            threadId,
            runResult.id,
            { tool_outputs },
          ),
        );
      }
    },
  );
}

View Example on GitHub