DeepResearch

An AI-powered multi-agent research assistant that turns a query into a structured, cited research report — with human-in-the-loop plan approval and real-time streaming.

2026-07-02.16-05-52.mp4

---

How It Works

1. User submits a research query with optional topic filters (News, Academic, Finance, Patents)
2. A Planner Agent generates a Problem Statement + up to 5 independent research sub-tasks
3. User reviews and approves the plan (or requests revisions in natural language)
4. Up to 5 Parallel Research Workers search the web simultaneously via Tavily
5. An Aggregator Agent synthesizes all findings into a structured markdown report
6. User can download the report as a .md file

---

Tech Stack

Layer	Technology
Agent Graph	LangGraph (HITL, Send() API, MemorySaver)
LLM	DeepSeek Pro (planner, aggregator) + Flash (approval, researchers)
Web Search	Tavily API
Backend	FastAPI + SSE streaming
Frontend	Vanilla JS + Nginx
Containerization	Docker + Docker Compose

---

Features

• Multi-agent graph — parallel researchers fan out via LangGraph's Send() API
• Human-in-the-Loop — review and revise the research plan before execution
• Real-time SSE streaming — live researcher progress updates in the UI
• Query validation — rejects vague queries before hitting the LLM
• Markdown export — download the final report as a .md file
• Session persistence — resume interrupted sessions via localStorage
• Dark/Light mode — theme toggle with persistent preference

---

Project Structure

research-bot/
├── backend/
│   ├── app/
│   │   ├── agents/          # planner, validator, approval, researcher, aggregator
│   │   ├── api/             # FastAPI routes + SSE
│   │   ├── graph/           # LangGraph state + builder
│   │   ├── tools/           # Tavily search tool
│   │   └── llm.py           # DeepSeek LLM config
│   └── Dockerfile
├── frontend/
│   ├── index.html
│   ├── app.js
│   └── style.css
├── docker-compose.yml
├── requirements.txt
└── .env.example

---

Getting Started

Prerequisites

• Docker + Docker Compose
• DeepSeek API key → platform.deepseek.com
• Tavily API key → tavily.com

1. Clone the repo

git clone https://github.com/Aryan-Pardeshi/deep-research.git
cd deep-research/research-bot

2. Set up environment variables

cp .env.example .env

Edit .env:

DEEPSEEK_API_KEY=your_key_here
DEEPSEEK_BASE_URL=https://api.deepseek.com
TAVILY_API_KEY=your_key_here

3. Run with Docker

docker compose up -d --build

• Frontend → http://localhost:80
• Backend → http://localhost:8000

---

Local Development (without Docker)

Backend

cd research-bot
pip install -r requirements.txt
uvicorn backend.app.main:app --reload --port 8000

Frontend

Open frontend/index.html directly in your browser, or serve with any static file server:

npx serve frontend

Make sure API_BASE_URL in app.js points to http://localhost:8000

---

API Endpoints

Method	Endpoint	Description
POST	/research/start	Validate query, generate PS + plan
POST	/research/approve	Resume graph, stream SSE events
GET	/research/result/{thread_id}	Fetch final report by thread

---

Author

Built by Aryan Pardeshi — open to AI/ML internship opportunities.

Connect: LinkedIn · GitHub