KEVIN 🔍

Knowledge Extraction and Verification Intelligent Network

A sophisticated LangGraph-based multi-agent deep research system that automates comprehensive research workflows through intelligent agent orchestration.

✨ Features

• Multi-Agent Architecture: Specialized agents for scoping, research, and report generation
• Intelligent Research Workflow: Automated research process from question clarification to final report
• Advanced Search Integration: Powered by Tavily search API with content summarization
• Flexible Model Support: Configurable OpenAI models through OpenRouter
• LangGraph Integration: Built on LangGraph for robust workflow orchestration
• TypeScript: Full type safety and modern development experience

🏗️ Architecture

KEVIN employs a multi-agent system with specialized roles:

🔍 Research Agents

• Scoping Agent: Clarifies research questions and defines scope
• Research Agent: Conducts deep research using search tools and reflection
• Multi-Agent Research: Coordinates multiple research workflows
• MCP Research Agent: Integrates with Model Context Protocol for enhanced capabilities

🧠 Core Components

• Search Tools: Tavily API integration with deduplication and summarization
• Reflection Tools: Strategic thinking and progress assessment
• Compression System: Intelligent research summarization
• Report Generation: Comprehensive final report compilation

🚀 Quick Start

Prerequisites

• Node.js 20+
• Yarn package manager
• OpenAI API key (via OpenRouter)
• Tavily Search API key

Installation

# Clone the repository
git clone <repository-url>
cd kevin-odr

# Install dependencies
yarn install

# Set up environment variables
cp .env.example .env
# Edit .env with your API keys

Environment Configuration

Create a .env file with the following variables:

OPENAI_API_KEY=your_openrouter_api_key
TAVILY_API_KEY=your_tavily_api_key

Running the System

# Build the project
yarn build

# Run tests
yarn test

# Run integration tests
yarn test:int

# Lint and format
yarn lint:all

📚 Usage

Using Individual Agents

The system provides several specialized agents accessible through the LangGraph configuration:

1. Scoping Agent

// Clarifies research questions and defines scope
import { agent as scopingAgent } from "./src/agents/scoping.js";

2. Research Agent

// Conducts focused research on specific topics
import { agent as researchAgent } from "./src/agents/research.js";

3. Full Research Pipeline

// Complete end-to-end research workflow
import { agent as fullAgent } from "./src/agents/full.js";

4. Multi-Agent Research

// Coordinates multiple research agents
import { agent as multiAgentResearch } from "./src/agents/research_multi_agent.js";

Example Research Workflow

// Initialize the research system
const state = {
  messages: [
    new HumanMessage("Research the top coffee shops in San Francisco"),
  ],
  research_brief: "",
  notes: [],
  final_report: "",
};

// Run the full research pipeline
const result = await fullAgent.invoke(state);
console.log(result.final_report);

🛠️ Development

Project Structure

src/
├── agents/          # Research agent implementations
│   ├── scoping.ts      # Question scoping and clarification
│   ├── research.ts     # Core research functionality
│   ├── research_mcp.ts # MCP integration
│   ├── research_multi_agent.ts # Multi-agent coordination
│   └── full.ts         # Complete pipeline
├── models.ts        # AI model configurations
├── tools.ts         # Research tools (search, reflection)
├── prompts.ts       # System prompts and templates
├── schemas.ts       # Data validation schemas
└── utils.ts         # Utility functions

⚙️ Configuration

Model Configuration

The system uses multiple specialized models configured in src/models.ts:

• Scoping Model: GPT-4o for question clarification
• Research Model: O4-mini for efficient research tasks
• Summarization Model: GPT-4o-mini for content summarization
• Compression Model: GPT-4.1-mini for research compression
• Supervisor Model: GPT-4.1 for agent coordination
• Writer Model: GPT-4.1 for report generation (32K context)

LangGraph Configuration

Agent graphs are defined in langgraph.json:

{
  "graphs": {
    "scoping": "./src/agents/scoping.ts:agent",
    "research": "./src/agents/research.ts:agent",
    "research_mcp": "./src/agents/research_mcp.ts:agent",
    "research_multi_agent": "./src/agents/research_multi_agent.ts:agent",
    "full": "./src/agents/full.ts:agent"
  }
}

🔧 Tools and Integrations

Search Tools

• Tavily Search: Web search with content summarization and deduplication
• Think Tool: Strategic reflection and decision-making support

External Services

• OpenRouter: Access to various OpenAI models
• Tavily API: Advanced web search capabilities
• Model Context Protocol (MCP): Extended tool integration

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Built with LangGraph for workflow orchestration
• Powered by Tavily for intelligent search
• Uses OpenRouter for model access

---

KEVIN - Intelligent Research, Automated