@llumiverse/drivers

Llumiverse - Universal LLM Connectors for Node.js

LLumiverse is a universal interface for interacting with Large Language Models, for the Typescript/Javascript ecosystem. It provides a lightweight modular library for interacting with various LLM models and execution platforms.

It solely focuses on abstracting LLMs and their execution platforms, and does not provide prompt templating, or RAG, or chains, letting you pick the best tool for the job.

The following LLM platforms are supported in the current version:

Provider	Completion	Chat	Model Listing	Fine-Tuning
OpenAI	✅	✅	✅	✅
Replicate	✅	✅	✅	✅
AWS Bedrock	✅	✅	✅	✅
HuggingFace Inference Endpoints	✅	✅	N/A	N/A
Google Vertex AI	✅	✅	Partial	In Progress
Together AI	✅	✅	✅	In Progress
Groq	✅	✅	✅	N/A

New capabilities and platform can easily be added by creating a new driver for the platform.

Requirements

• node v18+, or bun 1.0+

Installation

1. If you want to use llumiverse to execute prompt completion on various supported providers then install @llumiverse/core and @llumiverse/drivers

npm install @llumiverse/core @llumiverse/drivers

1. If you only want to use typescript types or other structures from llumiverse you only need to install @llumiverse/core

npm install @llumiverse/core

1. If you want to develop a new llumiverse driver for an unsupported LLM provider you only need to install @llumiverse/core

npm install @llumiverse/core

Usage

First, you need to instantiate a driver instance for the target LLM platform you want to interact too. Each driver accepts its own set of parameters when instantiating.

OpenAI driver

import { OpenAIDriver } from "@llumiverse/drivers";

// create an instance of the OpenAI driver 
const driver = new OpenAIDriver({
    apiKey: "YOUR_OPENAI_API_KEY"
});

Bedrock driver

In this example, we will instantiate the Bedrock driver using credentials from the Shared Credentials File (i.e. ~/.aws/credentials). Learn more on how to setup AWS credentials in node.

import { BedrockDriver } from "@llumiverse/drivers";

const driver = new BedrockDriver({
    region: 'us-west-2'
});

VertexAI driver

For the following example to work you need to define a GOOGLE_APPLICATION_CREDENTIALS environment variable.

import { VertexAIDriver } from "@llumiverse/drivers";
const driver = new VertexAIDriver({
    project: 'YOUR_GCLOUD_PROJECT',
    region: 'us-central1'
});

Replicate driver

import { ReplicateDriver } from "@llumiverse/drivers";

const driver = new ReplicateDriver({
    apiKey: "YOUR_REPLICATE_API_KEY"
});

TogetherAI driver

import { TogetherAIDriver } from "@llumiverse/drivers";

const driver = new TogetherAIDriver({
    apiKey: "YOUR_TOGETHER_AI_API_KEY"
});

HuggingFace driver

import { HuggingFaceIEDriver } from "@llumiverse/drivers";

const driver = new HuggingFaceIEDriver({
    apiKey: "YOUR_HUGGINGFACE_API_KEY",
    endpoint_url: "YOUR_HUGGINGFACE_ENDPOINT",
});

Listing available models

Once you instantiated a driver you can list the available models. Some drivers accept an argument for the listModel method to search for matching models. Some drivers like for example replicate are listing a preselected set of models. To list other models you need to perform a search by giving a text query as an argument.

In the following example, we are assuming that we have already instantiated a driver, which is available as the driver variable.

import { AIModel } from "@llumiverse/core";

// instantiate the desired driver
const driver = createDriverInstance();

// list available models on the target LLM. (some drivers may require a search parameter to discover more models)
const models: AIModel[] = await driver.listModels();

console.log('# Available Models:');
for (const model of models) {
    console.log(`${model.name} [${model.id}]`);
}

Execute a prompt

To execute a prompt we need to create a prompt in the LLumiverse format and pass it to the driver execute method.

The prompt format is very similar to the OpenAI prompt format. It is an array of messages with a content and a role property. The roles can be any of "user" | "system" | "assistant" | "safety".

The safety role is similar to system but has a greater precedence over the other messages. Thus, it will override any user or system message that is saying something contrary to the safety message.

In order to execute a prompt we also need to specify a target model, given a model ID which is known by the target LLM. We may also specify execution options like temperature, max_tokens etc.

In the following example, we are again assuming that we have already instantiated a driver, which is available as the driver variable.

Also, we are assuming the model ID we want to target is available as the model variable. To get a list of the existing models (and their IDs) you can list the model as we shown in the previous example

Here is an example of model IDs depending on the driver type:

• OpenAI: gpt-3.5-turbo
• Bedrock: arn:aws:bedrock:us-west-2::foundation-model/cohere.command-text-v14
• VertexAI: text-bison
• Replicate: meta/llama-2-70b-chat:02e509c789964a7ea8736978a43525956ef40397be9033abf9fd2badfe68c9e3
• TogetherAI: mistralai/Mistral-7B-instruct-v0.1
• HuggingFace: aws-mistral-7b-instruct-v0-1-015

import { PromptRole, PromptSegment } from "@llumiverse/core";


// instantiate the desired driver
const driver = createDriverInstance();
const model = "the-model-id"; // change with your desired model ID

// create the prompt.
const prompt: PromptSegment[] = [
    {
        role: PromptRole.user,
        content: 'Please, write a short story about winter in Paris, in no more than 512 characters.'
    }
]

// execute a model and wait for the response
console.log(`\n# Executing prompt on ${model} model: ${prompt}`);
const response = await driver.execute(prompt, {
    model,
    temperature: 0.6,
    max_tokens: 1024
});

console.log('\n# LLM response:', response.result)
console.log('# Response took', response.execution_time, 'ms')
console.log('# Token usage:', response.token_usage);

Execute a prompt in streaming mode

In this example, we will execute a prompt and will stream the result to display it on the console as it is returned by the target LLM platform.

Note that some models don't support streaming. In that case, the driver will simulate a streaming using a single chunk of text corresponding to the entire response.

import { PromptRole, PromptSegment } from "@llumiverse/core";

// instantiate the desired driver
const driver = createDriverInstance();
const model = "the-model-id"; // change with your desired model ID

// create the prompt.
const prompt: PromptSegment[] = [
    {
        role: PromptRole.user,
        content: 'Please, write a short story about winter in Paris, in no more than 512 characters.'
    }
]

// execute the prompt in streaming mode 
console.log(`\n# Executing prompt on model ${model} in streaming mode: ${prompt}`);
const stream = await driver.stream(prompt, {
    model,
    temperature: 0.6,
    max_tokens: 1024
});

// print the streaming response as it comes
for await (const chunk of stream) {
    process.stdout.write(chunk);
}

// when the response stream is consumed we can get the final reponse using stream.completion field.
const streamingResponse = stream.completion!;

console.log('\n# LLM response:', streamingResponse.result)
console.log('# Response took', streamingResponse.execution_time, 'ms')
console.log('# Token usage:', streamingResponse.token_usage);

Generate embeddings

LLumiverse drivers expose a method to generate vector embeddings for a given text. Drivers supporting embeddings as of v0.10.0 are bedrock, openai, vertexai. If embeddings are not yet supported the generateEmbeddings method will throws an error.

Here is an example on using the vertexai driver. For the example to work you need to define a GOOGLE_APPLICATION_CREDENTIALS env variable to be able to access your gcloud project

import { VertexAIDriver } from "@llumiverse/drivers";

const driver = new VertexAIDriver({
    project: 'your-project-id',
    region: 'us-central1' // your zone
});

const r = await vertex.generateEmbeddings({ content: "Hello world!" });

// print the vector
console.log('Embeddings: ', v.values);

The result object contains the vector as the values property, the model used to generate the embeddings and an optional token_count which if defined is the token count of the input text. Depending on the driver, the result may contain additional properties.

Also you can specify a specific model to be used or pass other driver supported parameter.

Example:

import { VertexAIDriver } from "@llumiverse/drivers";

const driver = new VertexAIDriver({
    project: 'your-project-id',
    region: 'us-central1' // your zone
});

const r = await vertex.generateEmbeddings({ 
    content: "Hello world!", 
    model: "textembedding-gecko@002",  
    task_type: "SEMANTIC_SIMILARITY"
});

// print the vector
console.log('Embeddings: ', v.values);

The task_type parameter is specific to the textembedding-gecko model.

Contributing

Contributions are welcome! Please see CONTRIBUTING.md for more details.

License

Llumivers is licensed under the Apache License 2.0. Feel free to use it accordingly.