JSPM

  • ESM via JSPM
  • ES Module Entrypoint
  • Export Map
  • Keywords
  • License
  • Repository URL
  • TypeScript Types
  • README
  • Created
  • Published
  • Downloads 135
  • Score
    100M100P100Q46527F
  • License MIT

Fetch streaming LLM responses as an async iterable

Package Exports

  • asyncllm
  • asyncllm/anthropic
  • asyncllm/gemini

Readme

asyncLLM

npm version License: MIT

Fetch LLM responses as an async iterable.

Features

  • 🚀 Lightweight (<2KB) and dependency-free
  • 🔄 Works with multiple LLM providers (OpenAI, Anthropic, Gemini, and more)
  • 🌐 Browser and Node.js compatible
  • 📦 Easy to use with ES modules

Installation

npm install asyncllm

Usage

Call asyncLLM() just like you would use fetch with any LLM provider with streaming responses.

The result is an async generator that yields objects with content, tool, and args properties.

For example, to update the DOM with the LLM's response:

<!doctype html>
<html lang="en">
  <body>
    <div id="output"></div>
  </body>

  <script type="module">
    import { asyncLLM } from "https://cdn.jsdelivr.net/npm/asyncllm@1";

    const apiKey = "YOUR_API_KEY";

    // Example usage with OpenAI
    for await (const { content } of asyncLLM("https://api.openai.com/v1/chat/completions", {
      method: "POST",
      headers: {
        "Content-Type": "application/json",
        Authorization: `Bearer ${apiKey}`,
      },
      body: JSON.stringify({
        model: "gpt-4o-mini-mini",
        // You MUST enable streaming, else the API will return an {error}
        stream: true,
        messages: [{ role: "user", content: "Hello, world!" }],
      }),
    })) {
      // Update the output in real time.
      document.getElementById("output").textContent = content;
    }
  </script>
</html>

Node.js or bundled projects

import { asyncLLM } from "asyncllm";

// Usage is the same as in the browser example

API

asyncLLM(request: string | Request, options?: RequestInit, config?: SSEConfig): AsyncGenerator<LLMEvent, void, unknown>

Fetches streaming responses from LLM providers and yields events.

  • request: The URL or Request object for the LLM API endpoint
  • options: Optional fetch options
  • config: Optional configuration object for SSE handling
    • fetch: Custom fetch implementation (defaults to global fetch)
    • onResponse: Async callback function that receives the Response object before streaming begins. If the callback returns a promise, it will be awaited before continuing the stream.

Returns an async generator that yields LLMEvent objects.

LLMEvent

  • content: The text content of the response
  • tools: Array of tool call objects with:
    • name: The name of the tool being called
    • args: The arguments for the tool call as a JSON-encoded string, e.g. {"order_id":"123456"}
    • id: Optional unique identifier for the tool call (e.g. OpenAI's call_F8YHCjnzrrTjfE4YSSpVW2Bc or Anthropic's toolu_01T1x1fJ34qAmk2tNTrN7Up6. Gemini does not return an id.)
  • message: The raw message object from the LLM provider (may include id, model, usage stats, etc.)
  • error: Error message if the request fails

Examples

OpenAI

import { asyncLLM } from "https://cdn.jsdelivr.net/npm/asyncllm@1";

const body = {
  model: "gpt-4o-mini",
  // You MUST enable streaming, else the API will return an {error}
  stream: true,
  messages: [{ role: "user", content: "Hello, world!" }],
};

for await (const data of asyncLLM("https://api.openai.com/v1/chat/completions", {
  method: "POST",
  headers: { "Content-Type": "application/json", Authorization: `Bearer ${apiKey}` },
  body: JSON.stringify(body),
})) {
  console.log(data);
}

This will log something like this on the console:

{ content: "", tool: undefined, args: undefined, message: { "id": "chatcmpl-...", ...} }
{ content: "Hello", tool: undefined, args: undefined, message: { "id": "chatcmpl-...", ...} }
{ content: "Hello!", tool: undefined, args: undefined, message: { "id": "chatcmpl-...", ...} }
{ content: "Hello! How", tool: undefined, args: undefined, message: { "id": "chatcmpl-...", ...} }
...
{ content: "Hello! How can I assist you today?", tool: undefined, args: undefined, message: { "id": "chatcmpl-...", ...} }

Anthropic

The package includes an Anthropic adapter that converts OpenAI-style requests to Anthropic's format, allowing you to use the same code structure across providers.

import { asyncLLM } from "https://cdn.jsdelivr.net/npm/asyncllm@1";
import { anthropic } from "https://cdn.jsdelivr.net/npm/asyncllm@1/dist/anthropic.js";

// You can use the anthropic() adapter to convert OpenAI-style requests to Anthropic's format.
const body = anthropic({
  // Same as OpenAI example above
});

// Or you can use the asyncLLM() function directly with the Anthropic API endpoint.
const body = {
  model: "claude-3-haiku-20240307",
  // You MUST enable streaming, else the API will return an {error}
  stream: true,
  max_tokens: 10,
  messages: [{ role: "user", content: "What is 2 + 2" }],
};

for await (const data of asyncLLM("https://api.anthropic.com/v1/messages", {
  headers: { "Content-Type": "application/json", "x-api-key": apiKey },
  body: JSON.stringify(body),
})) {
  console.log(data);
}

The Anthropic adapter supports:

  • System messages
  • Multi-modal content (text and images only, no audio support)
  • Model parameters (temperature, max_tokens, top_p, stop, metadata.user_id, but not n, presence_penalty, frequency_penalty, logprobs, top_logprobs)
  • User metadata
  • Function/tool calling with parallel execution control
  • Stop sequences

Gemini

The package includes a Gemini adapter that converts OpenAI-style requests to Gemini's format, allowing you to use the same code structure across providers.

import { asyncLLM } from "https://cdn.jsdelivr.net/npm/asyncllm@1";
import { gemini } from "https://cdn.jsdelivr.net/npm/asyncllm@1/dist/gemini.js";

// You can use the gemini() adapter to convert OpenAI-style requests to Gemini's format.
const body = gemini({
  // Same as OpenAI example above
});

// Or you can use the asyncLLM() function directly with the Gemini API endpoint.
const body = {
  contents: [{ role: "user", parts: [{ text: "What is 2+2?" }] }],
};

for await (const data of asyncLLM(
  // You MUST use a streaming endpoint, else the API will return an {error}
  "https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash-8b:streamGenerateContent?alt=sse",
  {
    method: "POST",
    headers: {
      "Content-Type": "application/json",
      Authorization: `Bearer ${apiKey}`,
    },
    body: JSON.stringify(body),
  },
)) {
  console.log(data);
}

The Gemini adapter supports:

  • System messages
  • Multi-modal content (text, images, audio via URL or data URI)
  • Model parameters (temperature, max_tokens, top_p, stop, n, presence_penalty, frequency_penalty, logprobs, top_logprobs, but not metadata)
  • Function calling (no parallel execution support)
  • JSON mode and schema validation
  • Stop sequences
  • Multiple candidates

Function Calling

asyncLLM supports function calling (aka tools). Here's an example with OpenAI:

for await (const { tools } of asyncLLM("https://api.openai.com/v1/chat/completions", {
  method: "POST",
  headers: {
    "Content-Type": "application/json",
    Authorization: `Bearer ${apiKey}`,
  },
  body: JSON.stringify({
    model: "gpt-4o-mini",
    stream: true,
    messages: [
      { role: "system", content: "Get delivery date for order" },
      { role: "user", content: "Order ID: 123456" },
    ],
    tool_choice: "required",
    tools: [
      {
        type: "function",
        function: {
          name: "get_delivery_date",
          parameters: { type: "object", properties: { order_id: { type: "string" } }, required: ["order_id"] },
        },
      },
    ],
  }),
})) {
  console.log(JSON.stringify(tools));
}

tools is an array of objects with name, id (for Anthropic and OpenAI, not Gemini), and args properties. It streams like this:

[{"name":"get_delivery_date","id":"call_F8YHCjnzrrTjfE4YSSpVW2Bc","args":""}]
[{"name":"get_delivery_date","id":"call_F8YHCjnzrrTjfE4YSSpVW2Bc","args":"{\""}]
[{"name":"get_delivery_date","id":"call_F8YHCjnzrrTjfE4YSSpVW2Bc","args":"{\"order"}]
[{"name":"get_delivery_date","id":"call_F8YHCjnzrrTjfE4YSSpVW2Bc","args":"{\"order_id"}]
[{"name":"get_delivery_date","id":"call_F8YHCjnzrrTjfE4YSSpVW2Bc","args":"{\"order_id\":\""}]
[{"name":"get_delivery_date","id":"call_F8YHCjnzrrTjfE4YSSpVW2Bc","args":"{\"order_id\":\"123"}]
[{"name":"get_delivery_date","id":"call_F8YHCjnzrrTjfE4YSSpVW2Bc","args":"{\"order_id\":\"123456"}]
[{"name":"get_delivery_date","id":"call_F8YHCjnzrrTjfE4YSSpVW2Bc","args":"{\"order_id\":\"123456\"}"}]

Use a library like partial-json to parse the args incrementally.

Config

asyncLLM accepts a config object with the following properties:

  • fetch: Custom fetch implementation (defaults to global fetch).
  • onResponse: Async callback function that receives the Response object before streaming begins. If the callback returns a promise, it will be awaited before continuing the stream.

Here's how you can use a custom fetch implementation:

import { asyncLLM } from "https://cdn.jsdelivr.net/npm/asyncllm@1";

const body = {
  // Same as OpenAI example above
};

// Optional configuration. You can ignore it for most use cases.
const config = {
  onResponse: async (response) => {
    console.log(response.status, response.headers);
  },
  // You can use a custom fetch implementation if needed
  fetch: fetch,
};

for await (const { content } of asyncLLM(
  "https://api.openai.com/v1/chat/completions",
  {
    method: "POST",
    headers: { "Content-Type": "application/json", Authorization: `Bearer ${apiKey}` },
    body: JSON.stringify(body),
  },
  config,
)) {
  console.log(content);
}

Streaming from text

You can parse streamed SSE events from a text string (e.g. from a cached response) using the provided fetchText helper:

import { asyncLLM } from "https://cdn.jsdelivr.net/npm/asyncllm@1";
import { fetchText } from "https://cdn.jsdelivr.net/npm/asyncsse@1/dist/fetchtext.js";

const text = `
data: {"candidates": [{"content": {"parts": [{"text": "2"}],"role": "model"}}]}

data: {"candidates": [{"content": {"parts": [{"text": " + 2 = 4\\n"}],"role": "model"}}]}

data: {"candidates": [{"content": {"parts": [{"text": ""}],"role": "model"}}]}
`;

// Stream events from text
for await (const event of asyncLLM(text, {}, { fetch: fetchText })) {
  console.log(event);
}

This outputs:

{ data: "Hello" }
{ data: "World" }

This is particularly useful for testing SSE parsing without making actual HTTP requests.

Error handling

If an error occurs, it will be yielded in the error property. For example:

for await (const { content, error } of asyncLLM("https://api.openai.com/v1/chat/completions", {
  method: "POST",
  // ...
})) {
  if (error) console.error(error);
  else console.log(content);
}

The error property is set if:

  • The underlying API (e.g. OpenAI, Anthropic, Gemini) returns an error in the response (e.g. error.message or message.error or error)
  • The fetch request fails (e.g. network error)
  • The response body cannot be parsed as JSON

Changelog

  • 2.1.0: Added id to tools to support unique tool call identifiers from providers
  • 2.0.1: Multiple tools support.
    • Breaking change: tool and args are not part of the response. Instead, it has tools, an array of { name, args }
    • Fixed Gemini adapter to return toolConfig instead of toolsConfig
  • 1.2.2: Added streaming from text documentation via config.fetch. Upgrade to asyncSSE 1.3.1 (bug fix).
  • 1.2.1: Added config.fetch for custom fetch implementation
  • 1.2.0: Added config.onResponse(response) that receives the Response object before streaming begins
  • 1.1.3: Ensure max_tokens for Anthropic. Improve error handling
  • 1.1.1: Added Anthropic adapter
  • 1.1.0: Added Gemini adapter
  • 1.0.0: Initial release with asyncLLM and LLMEvent

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the MIT License - see the LICENSE file for details.