AutoResearcher

A multi-agent pipeline for intelligent, automated web research.

A 4-agent research pipeline (search → read → write → critique) wrapped in a FastAPI backend that streams live progress over SSE, with a React + Tailwind glassmorphism frontend that visualizes the pipeline as it runs. Powered by Groq and Llama 3!

research-pipeline/
├── backend/        FastAPI app + your original LangChain agents/chains
│   ├── agents.py        (your agent/chain definitions — small bug fixes only)
│   ├── tools.py          (your web_search / scrape_url tools)
│   ├── pipeline.py       (your original CLI pipeline, unchanged in spirit)
│   ├── pipeline_stream.py (generator version of the pipeline — yields an event per step)
│   ├── main.py           (FastAPI app, exposes GET /api/research as an SSE stream)
│   └── requirements.txt
└── frontend/       Vite + React + Tailwind UI
    ├── src/components/   (AgentNode, PipelineFlow, ReportPanel, CriticPanel, etc.)
    └── src/App.jsx        (wires the UI to the SSE stream)

1. Backend setup

cd backend
python -m venv .venv
source .venv/bin/activate        # .venv\Scripts\activate on Windows
pip install -r requirements.txt
cp .env.example .env             # then fill in your real keys

Your .env needs:

GROQ_API_KEY=...
TAVILY_API_KEY=...

Run the API:

uvicorn main:app --reload --port 8000

Health check: GET http://localhost:8000/api/health → {"status": "ok"}

The original CLI pipeline still works unchanged: python pipeline.py.

2. Frontend setup

cd frontend
npm install
cp .env.example .env             # defaults to http://localhost:8000, change if needed
npm run dev

Open the printed local URL (usually http://localhost:5173). Type a topic, hit Run research, and watch the four agent nodes light up live as the backend streams each step.

How the streaming works

GET /api/research?topic=... is a Server-Sent Events stream. The blocking LangChain .invoke() calls run in a background thread so the FastAPI event loop stays responsive; each completed step gets pushed onto a queue and streamed to the browser as a small JSON event:

{"step": "search", "status": "complete", "data": "..."}

The frontend listens with the native EventSource API — no extra library needed for the streaming itself.

Notes / things fixed while wrapping this

• agents.py had load_dotenv without calling it (load_dotenv()), so the .env file was never actually being read. Fixed.
• tools.py had a module-level print(web_search.invoke(...)), which fired a real Tavily search every time the module was imported (including on every backend auto-reload). Removed.
• CORS is currently locked to localhost:5173 in main.py — update allow_origins there before deploying anywhere else.