Multi Agent Research Assistant

Full-stack multi-agent research assistant that turns a user query into a structured research report through a search, read, and critique workflow.

Overview

This project is built to reduce manual research time by orchestrating specialized agents instead of relying on a single model response.

The backend runs a three-stage pipeline:

1. Search agent finds relevant sources from the web.
2. Reader agent opens the most relevant pages and extracts usable content.
3. Critic agent refines the collected notes into a final report.

The frontend consumes streamed events over Server-Sent Events, so users can see each phase of the research process in real time.

Key Features

• Multi-agent workflow with separated responsibilities for search, reading, and critique
• Real-time streaming updates for phase status, tool calls, and final output
• Prompt-engineered agent pipeline powered by LangChain
• OpenRouter-backed LLM responses
• Tavily-powered web search for source discovery
• Webpage content extraction and cleaning before report generation
• Markdown report rendering with grouped citations
• Copy-to-clipboard support for the final report
• Stop and reset controls for long-running research sessions

Tech Stack

Layer	Tools	Purpose
Frontend	React, TypeScript, Vite	User interface and app shell
UI	Tailwind CSS, Motion, Lucide React	Styling, motion, and icons
Backend	FastAPI, Python	API layer and streaming endpoint
AI Orchestration	LangChain	Agent creation and prompt-driven workflow
LLM Provider	OpenRouter	Model access for agent responses
Search	Tavily	Web search and source discovery
Content Extraction	BeautifulSoup, Requests	Page reading and text cleaning
Markdown	React Markdown, remark-gfm	Report rendering

Architecture

The system is split into two parts:

Backend

• search_agent uses Tavily search to collect relevant sources.
• reader_agent fetches and cleans webpage content.
• critic_agent synthesizes the final report from the collected notes.
• research_pipeline_stream.py runs the pipeline on a background thread and emits structured events.
• app/main.py exposes the API and enables CORS for the frontend.

Frontend

• useResearch.ts manages the run state and handles incoming stream events.
• ResearchTimeline shows each phase and its tool activity.
• ReportView renders the final report in Markdown and supports citations.
• ChatInput submits queries and provides stop/reset behavior for active runs.

Environment Variables

Create .env files from the examples below.

Backend .env

Variable	Description
OPENROUTER_API_KEY	API key used for OpenRouter LLM access
TAVILY_API_KEY	API key used for web search
FRONTEND_URL	Optional frontend origin for CORS, for example the deployed app URL

Frontend .env

Variable	Description
VITE_API_BASE_URL	Backend API base URL, default is http://localhost:8000

Local Setup

1. Backend

cd backend
uv sync
cp .env.example .env
uv run uvicorn app.main:app --reload --port 8000

2. Frontend

cd frontend
npm install
cp .env.example .env
npm run dev

API Endpoints

Method	Endpoint	Description
GET	/api/health	Health check endpoint
POST	/api/research/stream	Starts a streamed research run

Research Flow

1. The user submits a topic or question.
2. The backend creates a shared research state object.
3. The search agent gathers candidate sources.
4. The reader agent extracts and cleans source content.
5. The critic agent composes the final report.
6. The frontend updates the timeline and renders the report as the stream progresses.

Project Structure

backend/
  app/
  agents/
  prompts/
  tools/
  research_pipeline.py
  research_pipeline_stream.py
frontend/
  src/
    components/
    hooks/
    lib/
    types/

• Live demo: https://researcharena.vercel.app

Notes

• The backend is designed around streaming output, not a single blocking response.
• The report is rendered in Markdown, which keeps the output readable and citation-friendly.
• The frontend can stop an active run and reset the workspace for a new query.

MIT Licence