Type Prompt 🎭

A powerful TypeScript library for LLM prompt templating, inspired by Banks

🚀 Features

• 📝 Template rendering - Create prompts using the Nunjucks templating engine
• 💬 Chat messages - Easily generate chat-based prompts for modern LLMs
• 🔧 Filters and extensions - Apply transformations to your prompt content
• ⚡ Caching - Efficiently render prompts by avoiding redundant processing
• 🛠️ Tool calling - First-class support for function calling in LLMs
• 🖼️ Vision support - Add images to prompts for multimodal models

📦 Installation

npm install type-prompt

🎯 Quick Start

Basic Prompt

import { Prompt } from "type-prompt";

const p = new Prompt("Write a 500-word blog post on {{ topic }}.");
console.log(p.text({ topic: "AI frameworks" }));

📚 Examples

Chat Messages

import { Prompt } from "type-prompt";

const p = new Prompt(`
{% chat role="system" %}
You are a {{ persona }}.
{% endchat %}

{% chat role="user" %}
Hello, how are you?
{% endchat %}
`);

const messages = p.chatMessages({ persona: "helpful assistant" });
// Output:
// [
//   { role: 'system', content: 'You are a helpful assistant.' },
//   { role: 'user', content: 'Hello, how are you?' }
// ]

Prompt Caching (for Anthropic)

import { Prompt } from "type-prompt";

const p = new Prompt(`
{% chat role="user" %}
Analyze this book:

{{ book | cache_control("ephemeral") }}

What is the title of this book? Only output the title.
{% endchat %}
`);

const messages = p.chatMessages({ book: "This is a short book!" });
// The book content will be wrapped in a special content block with cache_control

Function Calling

import { Prompt } from "type-prompt";

function getLaptopInfo() {
  /**
   * Get information about the user laptop.
   */
  return "MacBook Pro, macOS 12.3";
}

const p = new Prompt(`
{% chat role="user" %}
{{ query }}
{{ getLaptopInfo | tool }}
{% endchat %}
`);

const messages = p.chatMessages({
  query: "Can you guess the name of my laptop?",
  getLaptopInfo,
});
// The tool will be properly formatted for LLM function calling

�� API Reference

Prompt

The main class for creating and rendering prompts.

new Prompt(template: string, options?: {
  name?: string;
  version?: string;
  metadata?: Record<string, any>;
  canaryWord?: string;
  renderCache?: RenderCache;
})

Methods

• text(data?: Record<string, any>): Render the prompt as plain text
• chatMessages(data?: Record<string, any>): Render the prompt as an array of chat messages
• canaryLeaked(text: string): Check if a canary word has leaked

AsyncPrompt

An asynchronous version of the Prompt class with the same API but providing Promise-based methods.

new AsyncPrompt(template: string, options?: {
  name?: string;
  version?: string;
  metadata?: Record<string, any>;
  canaryWord?: string;
  renderCache?: RenderCache;
})

Methods

• text(data?: Record<string, any>): Returns a Promise that resolves to the rendered text
• chatMessages(data?: Record<string, any>): Returns a Promise that resolves to an array of chat messages
• canaryLeaked(text: string): Check if a canary word has leaked

Filters

• cache_control(text: string, cacheType: string = "ephemeral"): Mark text for caching
• image(source: string): Include an image in the prompt
• tool(function: Function): Convert a function to a tool for function calling

Extensions

• chat: Define a chat message block
• completion: Generate text using an LLM during template rendering

🤝 Contributing

Contributions, issues, and feature requests are welcome! Feel free to check the issues page.

📝 License

This project is MIT licensed.