AI Multi-Agent Research Assistant

A multi-agent research system built with LangGraph and LangChain. You give it a topic, and 4 specialized agents — Researcher, Analyst, Writer, and Reviewer — work through a stateful pipeline to produce a polished, self-reviewed report.

The frontend is React + Vite with a black/white design. The backend is FastAPI streaming real-time agent progress via SSE. LangSmith handles full trace monitoring. Supabase stores every report.

---

Architecture

System Overview

---

LangGraph Agent State Machine

---

How It Works

Each agent reads from and writes to a shared ResearchState. The pipeline is linear — no conditional loops in this version.

User Query → Researcher → Analyst → Writer → Reviewer → Final Report

Agent progress streams live to the frontend via Server-Sent Events (SSE). When complete, the report is saved to Supabase.

---

The 4 Agents

| Agent | Built with | Does | |-------|-----------|------| | Researcher | LangChain ReAct Agent + DuckDuckGo | Runs 4–6 web searches, compiles raw findings | | Analyst | LCEL chain (prompt → llm → parser) | Extracts key insights and confidence scores | | Writer | LCEL chain | Writes a 1000+ word structured Markdown report | | Reviewer | LCEL chain | Scores the report and delivers a final edited version |

---

Stack

| Layer | Tech | Why | |-------|------|-----| | LLM | OpenAI gpt-4o-mini | Cheap, fast, high quality | | Agent Framework | LangChain 0.3 | LCEL chains, ReAct agents | | Orchestration | LangGraph 0.2 | Stateful pipeline | | Monitoring | LangSmith | Full trace visibility per run | | Web Search | DuckDuckGo | Free, no API key | | Backend | FastAPI + SSE | Real-time streaming | | Frontend | React + Vite | Clean, fast | | Database | Supabase (Postgres) | Stores all reports | | Backend Deploy | Render | Free tier | | Frontend Deploy | Vercel | Free tier |

---

Project Structure

ai-research-assistant/
├── api/
│   └── main.py              # FastAPI app with SSE streaming
│
├── agents/
│   ├── researcher.py        # ReAct Agent + DuckDuckGo
│   ├── analyst.py           # LCEL chain
│   ├── writer.py            # LCEL chain
│   └── reviewer.py          # LCEL chain
│
├── tools/
│   ├── web_search.py        # DuckDuckGo wrapper
│   └── text_tools.py
│
├── frontend/                # React + Vite app
│   ├── src/
│   │   ├── App.jsx          # Main app (query, agents, report)
│   │   ├── api.js           # SSE fetch + history API
│   │   └── index.css        # Black/white design system
│   ├── package.json
│   ├── vite.config.js
│   └── vercel.json
│
├── docs/
│   ├── system-architecture.png
│   ├── langgraph-state-machine.png
│   └── supabase_schema.sql  # Run this in Supabase SQL editor
│
├── config.py                # Env vars + LangSmith setup
├── database.py              # Supabase client
├── orchestrator.py          # LangGraph pipeline (for CLI use)
├── main.py                  # CLI runner
├── render.yaml              # Render deployment config
├── requirements.txt
└── .env.example

---

Running Locally

1. Clone and install

git clone https://github.com/kushalsai-01/research-agent
cd research-agent
pip install -r requirements.txt

2. Set up environment variables

cp .env.example .env
# Then edit .env and fill in your keys

You need at minimum:

• OPENAI_API_KEY — from platform.openai.com

Optional but recommended:

• SUPABASE_URL + SUPABASE_KEY — from supabase.com (free)
• LANGCHAIN_API_KEY — from smith.langchain.com (free)

3. Start the backend

uvicorn api.main:app --reload --port 8000
# http://localhost:8000
# http://localhost:8000/health  ← check config status

4. Start the frontend

cd frontend
npm install
npm run dev
# http://localhost:5173

---

Supabase Setup

1. Create a free project at supabase.com
2. Go to SQL Editor and run the contents of docs/supabase_schema.sql
3. Copy your project URL and anon key into .env

CREATE TABLE IF NOT EXISTS reports (
  id           UUID PRIMARY KEY DEFAULT gen_random_uuid(),
  topic        TEXT NOT NULL,
  final_report TEXT,
  raw_research TEXT,
  analysis     TEXT,
  created_at   TIMESTAMPTZ DEFAULT now()
);

The app works without Supabase — reports just won't be saved to history.

---

LangSmith Setup

1. Create a free account at smith.langchain.com
2. Create a project called ai-research-assistant
3. Copy your API key into .env:

LANGCHAIN_TRACING_V2=true
LANGCHAIN_API_KEY=ls__…[redacted]
LANGCHAIN_PROJECT=ai-research-assistant

Every agent run will now appear in LangSmith with full input/output traces, latency, and token usage across all 4 agents.

---

Deployment

Backend → Render

1. Push this repo to GitHub
2. render.com → New Web Service → connect the repo
3. Render will auto-detect render.yaml — just add your env vars:
   • OPENAI_API_KEY
   • SUPABASE_URL
   • SUPABASE_KEY
   • LANGCHAIN_API_KEY
4. It'll deploy automatically on every push to main

Frontend → Vercel

1. vercel.com → Add New Project → import this repo
2. Set Root Directory to frontend
3. Add environment variable:
   • VITE_API_URL = your Render backend URL (e.g. https://your-app.onrender.com)
4. Deploy — Vercel handles the rest

Verify deployment

GET https://your-app.onrender.com/health

Returns:

{
  "status": "ok",
  "langsmith_tracing": true,
  "supabase_configured": true,
  "model": "gpt-4o-mini"
}

---

Cost

Everything runs on free tiers.

| Service | Cost | |---------|------| | OpenAI gpt-4o-mini | ~$0.01–0.05 per report | | Render | Free | | Vercel | Free | | Supabase | Free (500MB) | | LangSmith | Free (5K traces/month) |

---

License

MIT

---

Credits

• LangChain
• LangGraph
• LangSmith
• Supabase
• OpenAI