AgentFlow

A beautiful, fully local multi-agent RAG system that turns your documents into intelligent research reports.

Built from scratch - no paid APIs, no cloud credits, everything runs on your laptop.

✨ What is AgentFlow?

AgentFlow is a production-grade multi-agent AI workflow dashboard that lets you:

• Upload your notes, research papers, PDFs, or lecture notes
• Give any goal in plain English
• Watch a team of specialized AI agents (Researcher → Critic → Writer → Editor) collaborate in real-time
• Get a clean, well-structured final report with proper reasoning trace

It uses Retrieval-Augmented Generation (RAG) so the agents stay grounded in your documents instead of hallucinating.

Everything runs 100% locally using Ollama + LangGraph — zero cost, zero data leaves your laptop.

I built a real agentic system that demonstrates:

• Modern AI orchestration (LangGraph)
• Production-ready RAG pipeline
• Full-stack development with beautiful UI
• Local-first architecture (privacy + zero cost)

This project helped me deeply understand agentic workflows, streaming, state management, and what companies actually look for in GenAI roles.

Features

• Multi-Agent Collaboration — Researcher, Critic, Writer & Editor work together
• Real RAG — Upload documents → agents retrieve relevant context
• Live Streaming UI — Watch agents think in real-time
• Beautiful Collapsible Agent Trace — Clean card-style interface
• Final Synthesis — Highlighted, polished final answer with Copy & Download as PDF
• Error Handling — Friendly messages if Ollama is not running
• 100% Local & Private — No OpenAI, no API keys, no internet after setup
• Modern Tech Stack — Next.js 15 + TypeScript + Tailwind + shadcn/ui

Tech Stack

• Frontend: Next.js 15 (App Router), TypeScript, Tailwind CSS, shadcn/ui
• AI Orchestration: LangGraph.js + Ollama (local LLM)
• RAG: ChromaDB + nomic-embed-text embeddings
• LLM: Qwen2.5:7B (or any model you prefer via Ollama)
• Deployment: Docker-ready (one-command local setup)

How to Run Locally (Super Simple)

Prerequisites

• Node.js (v20 or above)
• Ollama installed and running (ollama serve)
• At least one model pulled: ollama pull qwen2.5:7b

Step-by-step Setup

1. Clone the repo

   git clone https://github.com/yourusername/agentflow.git
   cd agentflow

2. Install dependencies

   npm install

3. Make sure Ollama is running and model is ready

   ollama pull qwen2.5:7b

4. Start the development server

   npm run dev

5. Open http://localhost:3000 in your browser

Done! No environment variables, no API keys, no complicated setup.

🚀 How to Use

1. Upload Documents (optional but recommended): Drag & drop .txt, .md, or .pdf files.
2. Write your goal: Tell the agents exactly what you need them to analyze.
3. Trigger Workflow: Click submit and watch the agents work and stream their thoughts in real-time!
4. Export: Once finished, you can copy the final answer or download the full synthesis report as a PDF.

---

🏗️ Architecture Overview

All inter-agent communication happens through a stateful LangGraph workflow.

• Supervisor ➔ Decides which agent acts next based on the pipeline state.
• Researcher ➔ Uses RAG to retrieve relevant, grounded information from the embedded documents.
• Critic ➔ Checks quality, evaluates gaps, and actively reduces hallucinations.
• Writer ➔ Creates a highly structured response utilizing only the approved research.
• Editor ➔ Polishes the final output for formatting, grammar, and tone.

---

🧠 Challenges I Faced & What I Learned

• Handling proper streaming with LangGraph's custom stream mode.
• Making RAG reliable with local embeddings and vector stores.
• Building a beautiful, responsive UI that feels premium and eliminates infinite scrolling.
• Debugging agent loops and managing complex state across multiple LLM calls.
• Making the entire system feel "alive" with real-time UI updates and trace logs.

This project taught me more about modern AI engineering and systems architecture than any course.

---

🗺️ Future Improvements (Roadmap)

• Docker + docker-compose one-click setup
• Support for more complex file types (advanced PDF and image parsing)
• Memory between sessions (persistent knowledge base)
• Evaluation metrics dashboard
• Export to Markdown/Notion integrations

---

👨‍💻 About the Developer

Built by Arun Eswaran Pre-final year B.Tech CSE core student at SRM University

Passionate about AI agents, full-stack development, and building real-world products that solve complex problems.

• 🔗 LinkedIn: linkedin.com/in/arun-eswaran-dev
• 🐙 GitHub: github.com/arun-1312

If you're a recruiter or a fellow student and find this project interesting, feel free to connect or star the repo! ⭐

Made with ❤️ and vibe coding.