llm-decision

LLM-powered control flow functions for TypeScript/JavaScript.

Installation

npm install llm-decision ai @ai-sdk/openai

Quick Start

import { configureLLMDecision, llm_if, llm_switch } from 'llm-decision';
import { openai } from '@ai-sdk/openai';

// Configure globally
configureLLMDecision({
  model: openai('gpt-4o-mini'),
});

// Use llm_if just like a real if
if (await llm_if("The user seems happy based on: 'I love this!'")) {
  console.log('User is happy!');
} else {
  console.log('User is not happy.');
}

// Use llm_switch just like a real switch
switch (await llm_switch("Classify the sentiment of: 'I love this!'", ['positive', 'negative', 'neutral'] as const)) {
  case 'positive': console.log('Celebrate!'); break;
  case 'negative': console.log('Comfort.'); break;
  case 'neutral':  console.log('Acknowledge.'); break;
}

API

configureLLMDecision(config)

Configure global settings for all LLM decision functions.

configureLLMDecision({
  model: openai('gpt-4o-mini'), // Required: Language model from Vercel AI SDK
  temperature: 0,               // Optional: Default 0 for deterministic results
  maxRetries: 2,                // Optional: Default 2
});

llm_if(condition, options?)

Evaluates a natural language condition and returns a Promise<boolean>. Use it as the condition of a real if statement.

if (await llm_if("The user's message indicates frustration")) {
  routeToSupport();
} else {
  continueNormally();
}

Per-call model override:

if (await llm_if("Complex decision requiring better model", { model: openai('gpt-4o') })) {
  // ...
}

llm_switch(condition, cases, options?)

Selects from a list of cases based on a natural language condition and returns Promise<T>. Use it as the expression of a real switch statement.

switch (await llm_switch("Classify the urgency: 'Server is down!'", ['low', 'medium', 'high', 'critical'] as const)) {
  case 'critical': page(); break;
  case 'high':     alert(); break;
  case 'medium':   queue(); break;
  case 'low':      log();   break;
}

Provider Support

This package uses the Vercel AI SDK for provider-agnostic LLM calls. Install the provider package you need:

# OpenAI
npm install @ai-sdk/openai

# Anthropic
npm install @ai-sdk/anthropic

# Google
npm install @ai-sdk/google

Then configure with your preferred provider:

import { anthropic } from '@ai-sdk/anthropic';

configureLLMDecision({
  model: anthropic('claude-3-5-sonnet-20241022'),
});

TypeScript

Full TypeScript support with proper type inference:

// Return type is inferred as boolean
const isValid = await llm_if("Check if valid");

// Return type is inferred as 'a' | 'b' | 'c'
const choice = await llm_switch("Pick one", ['a', 'b', 'c'] as const);

License

MIT

llm-decision