Introduction

ToolRAG provides a seamless solution for using an unlimited number of function definitions with Large Language Models (LLMs), without worrying about context window limitations, costs, or performance degradation.

🌟 Key Features

• Unlimited Tool Definitions: Say goodbye to context window constraints. ToolRAG dynamically selects only the most relevant tools for each query.
• Semantic Tool Search: Uses vector embeddings to find the most contextually relevant tools for a given user query.
• Cost Optimization: Reduces token usage by only including the most relevant function definitions.
• Performance Improvement: Prevents performance degradation that occurs when overwhelming LLMs with too many function definitions.
• MCP Integration: Works with any Model Context Protocol (MCP) compliant servers, enabling access to a wide ecosystem of tools.
• OpenAI Compatible: Format tools as OpenAI function definitions for seamless integration.

🔍 How It Works

1. Tool Registration: ToolRAG connects to MCP servers and registers available tools.
2. Embedding Generation: Tool descriptions and parameters are embedded using vector embeddings (OpenAI or Google).
3. Query Analysis: When a user query comes in, ToolRAG finds the most relevant tools via semantic search.
4. Tool Execution: Execute selected tools against the appropriate MCP servers.

Installation

npm install @antl3x/toolrag
# or
yarn add @antl3x/toolrag
# or
pnpm add @antl3x/toolrag

🚀 Quick Start

import { ToolRAG } from "@antl3x/toolrag";
import OpenAI from "openai";

// Initialize ToolRAG with MCP servers
const toolRag = await ToolRAG.init({
  mcpServers: [
    "https://mcp.pipedream.net/token/google_calendar",
    "https://mcp.pipedream.net/token/stripe",
    // Add as many tool servers as you need!
  ],
});

const userQuery =
  "What events do I have tomorrow? Also, check my stripe balance.";

// Get relevant tools for a specific query
const client = new OpenAI();

const response = await client.responses.create({
  model: "gpt-4o",
  input: userQuery,
  tools: await toolRag.listTools(userQuery),
});

// Execute the function calls from the LLM response
for (const call of response.output.filter(
  (item) => item.type === "function_call"
)) {
  const result = await toolRag.callTool(call.name, JSON.parse(call.arguments));
  console.log(result);
}

🏗️ Architecture

ToolRAG uses a Retrieval-Augmented Generation (RAG) approach optimized for tools:

1. Storage: LibSQL database to store tool definitions and their vector embeddings
2. Retrieval: Cosine similarity search to find the most relevant tools
3. Execution: Direct integration with MCP servers for tool execution

👨‍💻 Use Cases

• Multi-tool AI Assistants: Build assistants that can access hundreds of APIs
• Enterprise Systems: Connect to internal tools and services without context limits
• AI Platforms: Provide a unified interface for tool discovery and execution

🔧 Configuration Options

ToolRAG offers flexible configuration options:

• Multiple embedding providers (OpenAI, Google)
• Customizable relevance thresholds
• Database configuration for persistence

📝 License

Apache License 2.0