OpenAI Queue

This library simplifies handling rate limits when using the OpenAI API. It provides the APIQueue class that manages API calls across all models based on tokens and requests per minute. Also, a higher-level ModelAPIQueue class is available for managing different API models each with their distinct rate limits.

The library automatically retries each request on encountering errors, with a 10-second delay, up to 5 attempts per request. Presently, the library naively retries all errors, but contributions are welcome for more sophisticated error handling.

While the code should function in a browser environment, due to the nested tiktoken dependency which uses wasm, we haven't added support for browser testing yet. Contributions for this are also welcome.

Usage

To use the library, create an APIQueue instance. To override the default rate limits, a customModelConfigs object can be passed to the constructor.

const customModelConfigs = {
    "gpt-4": {
        requestsPerMinute: 250,
        tokensPerMinute: 50000,
    },
};

const APIQueue = new APIQueue("your-api-key", customModelConfigs);

Then, make API calls using the request() method of the APIQueue object. The request() method dispatches the API call to the appropriate queue based on the model string of the request. The request object format matches that in the openai library.

const request = {
    model: "gpt-4",
    // Other parameters...
};

const response = await APIQueue.request(request);

Supported Models and Default Rate Limits

The library provides default rate limits for the following models:

• gpt-3.5-turbo: 3500 requests per minute and 90000 tokens per minute
• gpt-3.5-turbo-0301: 3500 requests per minute and 90000 tokens per minute
• gpt-4: 200 requests per minute and 40000 tokens per minute
• gpt-4-0314: 200 requests per minute and 40000 tokens per minute

Classes

ModelAPIQueue

Handles rate-limited API calls for a single model. It requires tokensPerMinute, requestsPerMinute, model, and apiKey parameters during construction. API calls are made using the request() method.

APIQueue

Manages API calls across different models, each with its unique rate limits. It internally creates a ModelAPIQueue instance for each model. API calls are made using the request() method, which dispatches the API call to the correct queue based on the request's model string.

OpenAI Agent Class

The Agent class provides a high-level interface to OpenAI's GPT-4 models, wrapping the process of making API requests and managing conversational context.

Basic Usage

Import the Agent class:

import Agent from "./agent";

Manually set a properly configured ModelAPIQueue on the Agent class:

import ModelAPIQueue from "./ModelAPIQueue";

Agent.api = new ModelAPIQueue(/*configuration*/);

Create a new agent using the create() static method:

const agent = Agent.create();

This method takes an optional config object, which can include any properties that an OpenAI CreateChatCompletionRequest would accept, excluding the messages property, and an extra head property.

To interact with the agent, call the agent as a function:

agent("Hello, agent!").then((newAgent) => {
    console.log(newAgent.content);
});

The agent() function makes an API request to OpenAI, then returns a new agent with the same config but an updated head representing the assistant's latest message. You can use this new agent to continue the conversation.

Additional Agent Features

• System Messages: Add system messages to the conversation using the system method. If the previous message was a system message, this will append to it instead of creating a new message.
• Extending an Agent: Create a new agent with extra or changed configuration using the extend method. This new agent will maintain the same conversation head, but with a different configuration.
• Other Methods: The Agent class includes several getter methods: head, content, and messages to access details about the agent and the conversation it maintains.
  • head: the full message object of the head.
  • content: the content string of the head message.
  • messages: the complete message history of the Agent.

This module is written in TypeScript and includes type definitions for all methods and configurations.

Roadmap

• [x] Node.js support
• [ ] Browser testing
• [ ] Non-chat model support