AI.js

AI.js is the easiest way to add AI text and image capabilities to your node applications:

await AI("the color of the sky is"); // blue

await AI.Image("a red rose"); // <image buffer: red rose>

await AI.Image.Concept("a red rose"); // <image buffer: a red rose in realist style, watercolor>

Under the hood AI.js seamlessly integrates all the best AI APIs:

• Text: OpenAI, Anthropic
• Image: Replicate, StabilityAI
• Music: Coming Soon
• Video: Coming Soon
• Embeddings: Coming Soon

We're constantly looking for new models and APIs to add. In a future update, AI.js will support local models like Stable Diffusion and LLaMA/Alpaca with the same easy-to-use interface.

AI.js is under heavy development and still subject to breaking changes.

Features

• Easy to use
• Best LLM chat completion models (gpt-3.5-turbo, gpt-4, claude-v1, claude-instant-v1)
• Best hosted image generation APIs (stable-diffusion-xl-beta-v2-2-2, many other Stable Diffusion models)
• Same interface across all services
• Streaming is easy as {stream: true}
• ai CLI interface in your shell
• MIT license

Installation

Install AI.js via npm

npm install @themaximalist/ai.js

Enable at least one service by setting its environment API_KEY

export OPENAI_API_KEY=sk-...
export ANTHROPIC_API_KEY=sk-ant-...
export STABILITY_API_KEY=sk-...
export REPLICATE_API_KEY=sk-....

Usage

The default interface is text. AI.js lets you send one-off requests or build up complex message histories with the Large Language Model (LLM).

const AI = require("@themaximalist/ai.js");
await AI("what is the codeword?"); // i don't know any codewords

const ai = new AI("the codeword is blue");
await ai.chat("what is the codeword?"); // blue

System and User Prompts

Giving the LLM a role can help improve performance, this can be done through system and user prompts.

await AI.system("I am HexBot—I generate beautiful hex color schemes", "autumn tree");
// #F2AF80 (soft peach), #E78C4D (burnt orange), #C86018 (rusty red), #5B3B0B (deep brown), #252527 (dark grey)

OpenAI has mentioned user prompts may be more strongly enforced than system prompts.

await AI.user("You are HexBot—you generate beautiful hex color schemes", "autumn tree");
// #F2AF80 (soft peach), #E78C4D (burnt orange), #C86018 (rusty red), #5B3B0B (deep brown), #252527 (dark grey)

Providing system and user roles can also be used with message history.

const ai = new AI();
ai.system("I am HexBot—I generate beautiful hex color schemes");
await ai.chat("autumn tree");
// #663300 (dark brown), #CC6600 (deep orange), #FF9900 (bright orange), #FFCC00 (golden yellow), #663399 (deep purple)
await ai.chat("make it lighter");
// #885533 (light brown), #FF9933 (pale orange), #FFCC66 ( light orange), #FFEE99 (pale yellow), #9966CC (light purple)

Message History

Chat history can build up over time, or you can initialize with an existing history.

await AI([
    { role: "user", content: "remember the secret codeword is blue" },
    { role: "user", content: "what is the secret codeword I just told you?" },
]); // blue

Or initialize with an existing history, and continue the conversation.

const ai = new AI([
    { role: "user", content: "remember the secret codeword is blue" },
    { role: "user", content: "what is the secret codeword I just told you?" },
]);
await ai.send(); // blue
await ai.chat("now the codeword is red");
await ai.chat("what is the codeword?"); // red

Note: Anthropic's message format is text-based and less convenient, so AI.js uses the OpenAI message format, and it's converted automatically on-the-fly for Anthropic.

Streaming

Streaming is as easy as passing {stream: true} as the second options parameter. A generator is returned that yields the completion tokens in real-time.

const stream = await AI("the color of the sky is", { stream: true });
for await (const message of stream) {
    process.stdout.write(message);
}

AI() Interface

The AI() interface is the same whether you're using new or await. Also options aim to be the same everywhere you can pass them, and use the most local scope possible, falling back to more global defaults and environment variables when needed.

await|new AI(
    string | array,
    options = {
        service: "openai", // openai, anthropic
        model: "gpt-3.5-turbo", // gpt-3.5-turbo, gpt-4, claude-v1, claude-instant-v1
        parser: null, // optional content parser or stream parser
        stream: false,
        context: AI.CONTEXT_FULL, // slices of message history can be sent, but by default send everything
    }
);

• Static Methods

  • AI.system(prompt, input, option=null) one-time use system prompt
  • AI.user(prompt, input, options=null) one-time use user prompt

• Instance Methods

  • ai.send(options=null) send chat completion request to network
  • ai.chat(content, options=null) add user message and send chat completion request to network
  • ai.user(content) add user message
  • ai.system(content) add system message
  • ai.assistant(content) add assistant message

• Instance Properties

  • ai.messages[] message history
  • ai.lastMessage last message

Image Generation

AI.js provides powerful image generation functions through StabilityAI and Replicate. To get started, make sure you've set the STABILITY_API_KEY or REPLICATE_API_KEY environment variable.

await AI.Image("a red rose"); // <image buffer: red rose>

AI.js also provides a concept generator—a way of using chat completion to generate a great image prompt. Using a concept generator can result in significantly better generated images.

await AI.Image.Concept("a red rose"); // <image buffer: a red rose in realist style, watercolor>

This hits your LLM provider and generates a complex image prompt before sending it off to the image generation service.

You can also create an AI.Image object to retrieve properties like generated_prompt from AI.Image.Concept.

const image = new AI.Image("a red rose");
await image.concept(); // <image buffer: a red rose in realist style, watercolor>
console.log(image.generated_prompt); // a red rose in the realist style, watercolor

AI.Image() Interface

The AI.Image() interface takes an input prompt and a few options.

await|new Image(
    string,
    options = {
        service: "stability", // stability, replicate
        model: "stable-diffusion-xl-beta-v2-2-2",
        concept_service: "openai", // openai, anthropic
        concept_model: "gpt-3.5-turbo", // gpt-3.5-turbo, gpt-4, claude-v1, claude-instant-v1
        concept_prompt: "I am Concept2Prompt. My task is to generate rich and dynamic scenes ...",
    }
);

• Static Methods
  • Image.Concept(string, options=null) generate a concept and send image completion request to network
• Instance Methods
  • image.send(options=null) send image completion request to network
  • image.concept(prompt, options=null) generate concept prompt and then send image completion request to network
• Instance Properties
  • image.prompt prompt supplied by user
  • image.generated_prompt prompt created from concept()
  • service image service
  • model image model
  • concept_model LLM concept model
  • concept_service LLM concept service
  • concept_prompt LLM concept prompt (uses default but can be overridden)

Environment Variables

AI.js supports configuration through environment variables.

Configure API keys

export OPENAI_API_KEY=sk-...
export ANTHROPIC_API_KEY=sk-ant-...
export STABILITY_API_KEY=sk-...
export REPLICATE_API_KEY=sk-....

Configure Service

export AI_SERVICE=openai
export AI_SERVICE=anthropic

export AI_IMAGE_SERVICE=stability
export AI_IMAGE_SERVICE=replicate

Configure Model

export AI_MODEL=gpt-3.5-turbo
export AI_MODEL=gpt-4
export AI_MODEL=claude-v1
export AI_MODEL=claude-instant-v1

export AI_IMAGE_MODEL=stable-diffusion-xl-beta-v2-2-2

AI.js will make some effort to make sure the right models are used with the right service if no defaults are set. So you can switch to anthropic and AI.js will automatically use claude-v1.

AI Command in your Shell

AI.js provides a handy ai command that can be invoked from your shell. This is an extremely convenient way to call models and services with the full power of AI.js. Access it globally by installing npm install @themaximalist/ai.js -g or setting up an nvm environment.

> ai the color of the sky is
blue

Messages are streamed back in real time.

You can also initiate a --chat to remember message history and continue your conversation. Ctrl-C to quit.

> ai remember the codeword is blue. say ok if you understand --chat
OK, I understand.
> what is the codeword?
The codeword is blue.

Model and service can be specified on the fly

> ai the color of the sky is --service anthropic --model claude-v1
blue

Or ai will fallback to $AI_SERVICE and $AI_MODEL environment variables.

> export AI_SERVICE=anthropic
> export AI_MODEL=claude-v1
> ai the color of the sky is
blue # claude-v1 response

Here's the help interface

> ai
Usage: ai [options] [input]

Simple AI interface to gpt-3.5-turbo, gpt-4 and Claude

Arguments:
  input                    Input to send to AI service

Options:
  -V, --version            output the version number
  -s, --service <service>  AI Service (default: "openai")
  -m, --model <model>      Completion Model (default: "gpt-3.5-turbo")
  -c, --chat               Chat Mode
  -h, --help               display help for command

Debug

AI.js and ai use the debug npm module with the ai.js namespace, so you can view debug logs by setting the DEBUG environment variable.

> DEBUG=ai.js:* ai the color of the sky is
# debug logs
blue
> export DEBUG=ai.js:*
> ai the color of the sky is
# debug logs
blue

Examples

View examples on how to use AI.js.

Projects

AI.js is currently used in the following projects:

• Infinity Arcade

Author

• The Maximalist
• @themaximal1st

License

MIT