cross-llm

Use LLM and Vector Embedding APIs on the web platform. Uses standard fetch() and thus runs everywhere, including in Service Workers.

🌟 Features

The most simple API to use LLMs. It can hardly be easier than 1 function call 😉

AI models currently supported:

• ✅ OpenAI: Any OpenAI LLM, including GPT-4 and newer models.
  • ✅ Promise-based
  • ✅ Streaming
  • ✅ Single message system prompt (instruct)
  • ✅ Multi-message prompt (chat)
  • ✅ Cost model
  • ✅ Text Embedding
• ✅ Anthropic: The whole Claude model-series, including Opus.
  • ✅ Promise-based
  • ✅ Streaming
  • ✅ Single message system prompt (instruct)
  • ✅ Multi-message prompt (chat)
  • ✅ Cost model
  • 〰️ Text Embedding (Anthropic doesn't provide embedding endpoints)
• ✅ Perplexity: All models supported.
  • ✅ Promise-based
  • ✅ Streaming
  • ✅ Single message system prompt (instruct)
  • ✅ Multi-message prompt (chat)
  • ✅ Cost model (including flat fee)
  • 〰️ Text Embedding (Perplexity doesn't provide embedding endpoints)
• ✅ VoyageAI: Text Embedding models
  • ✅ Text Embedding
• ✅ Mixedbread AI: Text Embedding models, specifically for German
  • ✅ Text Embedding

AI providers and models to be supported soon:

• ❌ Google: The whole Gemeni model-series, including 1.5 Pro, Advanced.
• ❌ Cohere: The whole Command model-series, including Command R Plus.
• ❌ Ollama: All Ollama LLMs, including Llama 3.
• ❌ HuggingFace: All HuggingFace LLMs.

📚 Usage

1. 🔨 First install the library: npm/pnpm/yarn/bun install cross-llm

2. 💡 Take a look at the super-simple code examples.

Single System Prompt

import { systemPrompt } from "cross-llm";

const promptResonse = await systemPrompt("Respond with JSON: { works: true }", "anthropic", {
  model: "claude-3-haiku-20240307",
  temperature: 0.7,
  max_tokens: 4096
}, { apiKey: import.meta.env[`anthropic_api_key`] });

// promptResponse.message => {\n  "works": true\n}
// promptResponse.usage.outputTokens => 12
// promptResponse.usage.inputTokens => 42
// promptResponse.usage.totalTokens => 54
// promptResponse.price.input => 0.0000105
// promptResponse.price.output => 0.000015
// promptResponse.price.total => 0.0000255
// promptResponse.finishReason => "end_turn"
// promptResponse.elapsedMs => 888 // milliseconds elapsed
// promptResponse.raw => provider's raw completion response object, no mapping
// promptResponse.rawBody => the exact body object passed to the provider's completion endpoint

Text Embedding

import { embed } from "cross-llm";

const textEmbedding = await embed(["Let's have fun with JSON, shall we?"], "voyageai", {
  model: "voyage-large-2-instruct",
}, { apiKey: import.meta.env[`voyageai_api_key`], });

// textEmbedding.data[0].embedding => [0.1134245, ...] // n-dimensional embedding vector
// textEmbedding.data[0].index => 0
// textEmbedding.usage.totalTokens => 23
// textEmbedding.price.total => calculated price
// textEmbedding.elapsedMs => 564 // in milliseconds

Multi-Message Prompt, Streaming

import { promptStreaming, type PromptFinishReason, type Usage, type Price } from "cross-llm";

await promptStreaming(
  [
    {
      role: "user",
      content: "Let's have fun with JSON, shall we?",
    },
    {
      role: "assistant",
      content: "Yeah. Let's have fun with JSON.",
    },
    {
      role: "user",
      content: "Respond with JSON: { works: true }",
    },
  ],
  "openai",
  async (partialText: string, elapsedMs: number) => {
    // onChunk

    // stream-write to terminal
    process.stdout.write(partialText);
  },
  async (fullText: string, 
    elapsedMs: number,
    usage: Usage,
    finishReason: PromptFinishReason,
    price: Price) => {

    // onStop
    console.log("")
    console.log("parsed JSON", JSON.parse(fullText));
    console.log("finishReason", finishReason);
    console.log("elapsedMs", elapsedMs);
    console.log("usage", usage);
    console.log("price", price);
  },
  async (error: unknown, elapsedMs: number) => {
    // onError
    console.log("error", error, elapsedMs, 'ms elapsed');
  },
  {
    model: "gpt-4-turbo",
    temperature: 0.7,
    response_format: {
      type: "json_object",
    }
  },
  {
    // union of options passed down, mapped internally
    apiKey: import.meta.env[`openai_api_key`],
  },
);

3. 📋 Copy & Paste -> enjoy! 🎉

🔥 Contributing

Simply create an issue or fork this repository, clone it and create a Pull Request (PR). I'm just implementing the features, AI model providers, cost model mappings that I need, but feel free to simply add your models or implement new AI providers. Every contribution is very welcome! 🤗

List/verify supported models

Please verify that your model/provider has been added correctly in ./src/models.

npm run print-models

Write and verify example code

Please add example code for when you implement a new AI provider in ./examples.

npm run example openai.ts

or

npm run example voyageai-embedding.ts

Write tests for new AI providers

Please write and run unit/integration/e2e tests using jest by creating ./src/*.spec.ts test suites:

npm run test

Build a release

Run the following command to update the ./dist files:

npm run build

Create a new NPM release build:

npm pack

Check the package contents for integrity.

npm publish