Connect to and stream from any OpenAI/Anthropic API. Lightweight, high performance, and simple, thoughtful API made for developers. Tested on OpenAI, Ollama, and Claude.

npm install poon-llm

OpenAI Example

const llm = new LLM({
    'apiBase': 'https://api.openai.com',
    'secretKey': 'key',
    'model': 'gpt-4o',
    'systemPrompt': 'You are a helpful assistant.',
});

const response = await llm.chat('Why is the sky blue?');

Anthropic Example

const llm = new LLM({
    'protocol': 'anthropic',
    'apiBase': 'https://api.anthropic.com/v1/messages',
    'secretKey': 'key',
    'headers': {'Anthropic-Version': '2023-06-01'},
    'model': 'claude-3-opus-20240229',
    'systemPrompt': 'You are a helpful assistant.',
});

const response = await llm.chat('Why is the sky blue?');

Streaming

Streaming events occur at a fast rate, so to avoid crashing your server, poon-llm employs an efficient method to combat this: While an async onUpdate is executing, any chunks that come in will be ignored so that onUpdate will only be called as fast as your code can handle it. For example, if you are on a shared database that takes 1 second to write, your callbacks will fire back to back, after each write, and then once more at the very end.

const response = await llm.chat('Why is the sky blue?', {
    'onUpdate': text => Drafts.updateAsync({'_id': id}, {
        $set: {'body': text}
    }),
});

Chat: Other Options

| Option | Description | |---------------|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | json | Enable JSON output: Requests underlying LLM API to respond in JSON, also JSON-parses and returns response. You must request the reply to be in JSON form in the system prompt. An error message will appear if the word JSON is not detected in the prompt. | | onUpdate | Callback function that is called every time the model has more chunks to append to the response. | | context | Chat history for the conversation, must be an array of objects like {'role': String ('user' or 'assistant'), 'content': String}. | | temperature | Float value controlling randomness in boltzmann sampling. Lower is less random, higher is more random. | | maxTokens | Integer value controlling the maximum number of tokens generated. |

Hot tip for Chain of Thought Prompts

Although JSON is an option, it turns out that XML is generally a better choice for prompts with Chain of Thought, because the LLM has an easier time formatting it correctly, as it just needs to understand delimiters, rather than strict adherence to a certain syntax, and parsing is not too much harder. Here is an example showing how this can be done.

import { parseFromString } from 'dom-parser';

const response = await llm.chat(chatString, {
    'prefill': '<scratchpad>',
    'maxTokens': 2048,
});

const dom = parseFromString(response);
const [scratchpadNode] = dom.getElementsByTagName('scratchpad');
const [replyNode] = dom.getElementsByTagName('reply');
return {
    'scratchpad': scratchpadNode?.textContent.trim(),
    'body': replyNode?.textContent.trim(),
};