anubhavgupta/prompt-compose

A JS library that provides basic axioms for building and managing GPT prompts. No dependencies.

prompt-compose

This library provides basic axioms for building and managing GPT prompts. It helps in building small and reusable prompt components which can be composed together to build larger ones.

Installation:

NodeJS:

npm install prompt-compose

Browser:

// via CDL link
import { Statement } from "https://unpkg.com/[email protected]/dist/module.js";
const prompt = Statement("Hello World");

Basic axioms for building prompts:

Statement

Creates a simple line of text. It is a basic building block for building larger prompts.

import { Statement } from "prompt-compose";

const question = Statement("How to make a website?");
question();
// "How to make a website?"

Instruction

Creates a new Instruction of text with title and content.

import { Instruction } from "prompt-compose";

const problemInstruction = Instruction(
    "Problem Statement",
    "How to make a website?"
);

problemInstruction();

// Problem Statement:
//     How to make a website?

Instead of using string directly, Statement can also be used whithout any change in result. This is true for rest of the examples as well.

const problemInstruction = Instruction(
    Statement("Problem Statement"),
    Statement("How to make a website?")
);

problemInstruction();

// Problem Statement:
//     How to make a website?

Section

Creates a new Section of text with title and content. It is built using Instruction and adds few capabilities over and above it.

Extra capabilities:

1. Multiple individual contents can be passed content without using compose.

   Section("Main", "Instruction 1", "Instruction 2",..)`;

   Note: It uses compose internally.
2. Adds a divider at the end of the section.

import { Section, Instruction } from "prompt-compose";

const problemSection = Section(
    "Problem Statement",
    "How to make a website?"
);

// Problem Statement:
//     How to make a website?

// Section can be used to compose multiple Instructions as well.
const userSection = Section(
    "User",
    Instruction(
        "Problem",
        "How to build a website?"
    ),
    Instruction(
        "Format",
        "Answer should be in step wise format."
    ),
    
);

// User:
//     Problem:
//         How to build a website?

//     Format:
//         Answer should be in step wise format.

Format

Asks for a data structure and creates a prompt where it asks GPT to return the response in the given format.

import { Format } from "prompt-compose";

const formatTheReponse = Format([
    {
        stepIndex: "<number>",
        stepDetails: "<string>"
    }
]);

formatTheReponse();
// `Answer should be in a valid JSON, use the following structure: 
//     ResponseJSON: [{"stepIndex":"<number>","stepDetails":"<string>"}]`

System

Creates a system Instruction.

import { System } from "prompt-compose";

const systemStatement = System(
`You are an AI assistant, you help users solve the given problem. 
- Your answeres are brief and to the point.
- You only answer in the format given by the user.
- If you don't know the answer to a question, you truthfully say that I don't know it. 
- You can end your answer with "END_END_END" text.`
);

systemStatement();

// `System:
//     You are an AI assistant, you help users solve the given problem. 
//     - Your answeres are brief and to the point.
//     - You only answer in the format given by the user.
//     - If you don't know the answer to a question, you truthfully say that I don't know it. 
//     - You can end your answer with "END_END_END" text.`

compose

Let you compose multiple prompts together.

import { compose } from "prompt-compose";

// Example 1: 
const makePrompt = compose(
    systemStatement,
    problemSection,
);

makePrompt();

// System:
//     You are an AI assistant, you help users solve the given problem. 
//     - Your answeres are brief and to the point.
//     - You only answer in the format given by the user.
//     - If you don't know the answer to a question, you truthfully say that I don't know it. 
//     - You can end your answer with "END_END_END" text.
 
// Problem Statement:
//     How to make a website?
//
//

// Example 2: 
const makePrompt2 = compose(
    systemStatement,
    problemSection,
);

makePrompt2(
    "Solution:" // can also pass Statement("Solution:")
);

// Problem Statement:
//     How to make a website?
//
// Solution:

Divider

Configurable, adds a division between 2 given prompts. use setConfig() to configure the look and feel of Divider.

import { Divider, Statement } from "prompt-compose";

// Example 1: 
const promptWithDivider = compose(
    Statement("Do task 1."),
    Divider(),
    Statement("Do task 2.")
);
promptWithDivider()

// Do task 1.
// 
// Do task 2.


// Example 2: 
setConfig({
    dividerType: "-",
})

const promptWithDivider2 = compose(
    Statement("Do task 1."),
    Divider(),
    Statement("Do task 2.")
);
promptWithDivider2()

// Do task 1.
// -----
// Do task 2.
 

None

An empty input.

import { None, Section, Statement } from "prompt-compose";

const expectedResponse = Section(
        Statement("AI"),
        None()
);
expectedResponse();

// AI:
//   

Configuration

This lib provides getConfig() and setConfig() functions to configure various aspects of the Axioms.

import { setConfig, getConfig } from "prompt-compose";

setConfig({
    tabType: " ", // \t or space; String should be used for tab ?
    tabRepeat: 4, // number of times the above needs to be repeated ?
    dividerType: "-", // which string should be used as a divider
    dividerRepeat: 5, // number of times the above needs to be repeated ?
    dividerBefore: "\n", // string before the divider
    dividerAfter: "\n", // string after the divider
    instructionSpacing:"\n" // string to create space between the title and the content in an Instruction.
});

const config = getConfig(); // returns the config object.

// below is the default configuration:
{
    dividerType: " ",
    dividerRepeat: 5,
    dividerBefore: "\n",
    dividerAfter: "\n",
    tabType:" ",
    tabRepeat: 4,
    instructionSpacing:"\n"
}

Example code:

import { 
    System,
    instruction,
    Section,
    Format,
    compose,
    Statement,
    None,
    setConfig,
} from "prompt-compose";
import type { PromptIO } from "prompt-compose";

// set the desired configuration if required.
setConfig({
    tabType: " ",
    tabRepeat: 4,
    dividerType: "",
    dividerRepeat: 0,
    dividerBefore: "\n",
    dividerAfter: "\n",
    instructionSpacing: "\n"
});

// create a System prompt telling GPT how to act.
const systemPrompt = () => System(
`You are an AI assistant, you help users solve the given problem. 
- Your answeres are brief and to the point.
- You only answer in the format given by the user.
- If you don't know the answer to a question, you truthfully say that I don't know it. 
- You can end your answer with "END_END_END" text.`
);

// create a user prompt telling GPT what user is asking for and how they want the response to be like.
const userPrompt = (problemToSolve: PromptIO, responseFormat: Array<any> | Object)=>{
    return Section(
        "User",
        Instruction(
            "Format",
            Format(responseFormat)
        ),
        problemInstruction(problemToSolve)
    );
};

// create a AI response completion prompt.
// accepts a response Statement, fallsback to None(empty string)
const aiPrompt = (response: PromptIO = None())=>{
    return Instruction(
        "AI",
        response
    );
}

// Defines a new Instruction where user can pass their ask
// Used by userPrompt.
const problemInstruction = (problemToSolve: PromptIO) => Instruction(
    "Problem",
    problemToSolve
);

// format in which we need the GPT too respond.
const responseFormat = [
    {
        stepIndex: "<number>",
        stepDetails: "<string>"
    }
];

// Now we compose the above prompts to generate the desired prompt.
const getProblemSolver = (problemToSolve: PromptIO, responseFormat: Array<any> | Object)=>{
    return compose(
        systemPrompt(), 
        userPrompt(problemToSolve, responseFormat), // problemInstruction and responseFormat are passed here
        aiPrompt()
    );
}

const generateSolverPrompt = getProblemSolver("How to build a website?", responseFormat);
const promptGenerated = generateSolverPrompt();


// System:
//     You are an AI assistant, you help users solve the given problem. 
//     - Your answeres are brief and to the point.
//     - You only answer in the format given by the user.
//     - If you don't know the answer to a question, you truthfully say that I don't know it. 
//     - You can end your answer with "END_END_END" text.

// User:
//     Format:
//         Answer should be in a valid JSON, use the following structure:
//             ResponseJSON: [{"stepIndex":"<number>","stepDetails":"<string>"}]
    
//     Problem:
//         How to build a website?

// AI:
//    

const generateSolverPrompt = getProblemSolver("How to make a pizza?", [
    {
        stepIndex: "<number>",
        stepDetails: "<string>",
        requiresOven: "<boolean>"
    }
]);
const promptGenerated = generateSolverPrompt();


// System:
//     You are an AI assistant, you help users solve the given problem. 
//     - Your answeres are brief and to the point.
//     - You only answer in the format given by the user.
//     - If you don't know the answer to a question, you truthfully say that I don't know it. 
//     - You can end your answer with "END_END_END" text.
//
// User:
//     Format:
//         Answer should be in a valid JSON, use the following structure:
//             ResponseJSON: [{"stepIndex":"<number>","stepDetails":"<string>","requiresOven":"<boolean>"}]
//    
//     Problem:
//         How to make a pizza?
//
// AI:
//

Development

• All axioms like Statement, Instruction, etc. are in ./src/base-components directory.
• Configuration utilities are in ./src/config/
• Utilities like compose and Tabs are in ./src/utils/
• Tests are in ./tests/index.ts

# make required code changes then run following command to run tests
npm run test

# to build the project
npm run build

prompt-compose's Issues

Add centralized configuration support

Added Centralized configuration support
Centralized configuration support added for following components:
Divider
Tab
Instruction

Add development guide

Add a custom Test runner

Add Typescript support

Add a separate build for test

New target: Tests for running tests
Code Watch support
Migrated Tests to typescript