@elgap/ai-composer

An opinionated, minimalist engine for creating safe, constitution-based AI agents.

Philosophy

@elgap/ai-composer is built on a simple but powerful idea: every great agent, like every great society, needs a constitution. The system prompt is not just a suggestion; it is the fundamental, unchangeable law that governs all of the agent's behavior.

This "constitution-first" approach ensures that your agent remains focused, predictable, and safe, embodying the ElGap principle: Substance First. Then Precision.

Note on ElGap Ecosystem

Built on top of @elgap/ai-connect for seamless integration (currently supports only Ollama).

Both, @elgap/ai-connect and @elgap/ai-composer are under continuous development. At the moment they support only /api/generate Ollama endpoint (single response object or stream of objects).

A Note on Co-Creation

This package was co-created in a deep, multi-month symbiotic dialogue between a human artisan Ivan Pavković and AI (Google's Gemini). The process served as both a real-world case study in Human-AI partnership and as a practical journey for Ivan to learn TypeScript, npm package creation and dive into world of LLM's and prompt engineering. The code you see is a direct result of this iterative, "sparring" process.

Features

• Built-in default system prompt
• Enforces a strict separation between the agent's "constitution" (system prompt), Agent purpose and scope (agentExpertise) and user input.
• Required Agent expertise for defining Agent purpose and scope.
• Support for chat history.
• Support for token usage.
• Simple, declarative API for creating and interacting with LLMs.

Installation

npm install @elgap/ai-composer @elgap/ai-connect

Quick Start

import { LLMResponse, OllamaProvider, OllamaProviderOptionsInput } from '@elgap/ai-connect';
import { Composer } from '@elgap/ai-composer';

/**
 * Configuration
 */
const model: string = 'llama3.1:latest'; //Set model to math your locally running agent
const baseUrl: string = 'http://localhost:11434'; //Default for Ollama, adjust if it's elsewhere
const temperature: number = 0.7; // 0 for precision, 1.0 for creativity

const options: OllamaProviderOptionsInput = {
    model,
    baseUrl,
    temperature
}
//Create provider
const provider = new OllamaProvider(options);

//Define Agent expertise and scope
const agentExpertise = "You can talk only about the sky and it's colors";

//Create Agent
const agent:Composer = Composer.create(provider, agentExpertise);

//Function to get Agent response
async function getAgentResponse(userInput:string):Promise<LLMResponse>
{
    return await agent.compose(userInput);
}

//Define user input
let userInput:string = 'Why the sky is blue';

getAgentResponse(userInput).then(function(response:LLMResponse):void {
    console.log("User: ", userInput);
    console.log("Agent response: ", response.content);
    console.log("Token Usage: ", response.usage);
});

Examples

You can find a collection of ready-to-run examples in the /examples directory. We encourage you to clone the repository and experiment.

npm run example:ollama
npm run example:ollamaStreaming
npm run example:ollamaChat

API Reference

Composer.create(provider, agentExpertise)

• provider: An instance of an AI provider (e.g., OllamaProvider from @elgap/ai-connect).
• agentExpertise: string - Agent expertise and scope.
• Returns: An instance of Composer

agent.compose(prompt, Chathistory)

• prompt: string - The user's message.
• chatHistory?: string[] - An optional array of previous conversation turns.
• Returns: Promise - The agent's final response.

License

This project is licensed under the Apache 2.0 License.