🔬 Deep Research AI Agent

An AI-powered research assistant that combines natural language processing with real-time web scraping to provide comprehensive, research-backed answers to your questions.

Features

• Natural Language Queries - Ask questions in plain English
• Intelligent Search Detection - Automatically determines when web search is needed
• Real-Time Web Scraping - Fetches current information via BrightData SERP API
• Self-Hosted LLM - Uses Nosana's decentralized GPU infrastructure for AI inference
• Beautiful Response Formatting - Renders markdown with tables, links, and rich formatting

Architecture

┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐
│   React Frontend │────▶│  FastAPI Backend │────▶│   Nosana LLM    │
│   (Vite + TS)   │     │   (Python)       │     │   (Ollama API)  │
└─────────────────┘     └────────┬─────────┘     └─────────────────┘
                                 │
                                 ▼
                        ┌─────────────────┐
                        │   BrightData    │
                        │   (Web Search)  │
                        └─────────────────┘

Tech Stack

Backend

• FastAPI - Python web framework
• httpx - Async HTTP client
• Pydantic - Data validation

Frontend

• React 19 - UI framework
• TypeScript - Type safety
• Vite - Build tool
• react-markdown - Markdown rendering with GFM support

External Services

• Nosana - Decentralized GPU compute for LLM inference
• BrightData - Web scraping and SERP API

Getting Started

Prerequisites

• Python 3.10+
• Node.js 18+
• Nosana endpoint URL
• BrightData API token

Installation

1. Clone the repository

   git clone <repository-url>
   cd deep_research_agent

2. Install all dependencies

   make install

   Or install separately:

   # Backend
   cd services/backend && pip3 install -r requirements.txt
   
   # Frontend
   cd services/frontend && npm install

3. Configure environment variables

   Create a .env file in services/backend/:

   NOSANA_OLLAMA_URL=your-nosana-endpoint-here
   NOSANA_MODEL=gpt-oss:20b
   BRIGHTDATA_API_TOKEN=your-brightdata-token-here
   PORT=8000

Running the Application

Start both services:

make dev

Or run them separately:

Backend (runs on http://localhost:8000):

make dev-backend

Frontend (runs on http://localhost:5173):

make dev-frontend

API Endpoints

Endpoint	Method	Description
/	GET	Health check - returns API status
/health	GET	Health check endpoint
/api/query	POST	Submit a query for research

Query Request

POST /api/query
{
  "query": "What are the latest developments in AI?"
}

Query Response

{
  "answer": "Based on my research, here are the latest developments..."
}

How It Works

1. Query Analysis - The LLM determines if the question requires real-time web data
2. Search Query Extraction - Extracts optimal search terms from the user's question
3. Web Scraping - BrightData fetches Google search results in real-time
4. Answer Generation - The LLM synthesizes search results into a comprehensive answer

Project Structure

deep_research_agent/
├── services/
│   ├── backend/
│   │   ├── main.py          # FastAPI application entry point
│   │   ├── agent.py         # WebSearchAgent orchestration logic
│   │   ├── llm.py           # Nosana LLM client (OpenAI-compatible)
│   │   ├── scraper.py       # BrightData web scraping client
│   │   ├── requirements.txt # Python dependencies
│   │   └── .env.example     # Environment template
│   └── frontend/
│       ├── src/
│       │   ├── App.tsx      # Main React component
│       │   └── App.css      # Styling with markdown support
│       ├── package.json     # Node dependencies
│       └── vite.config.ts   # Vite configuration
├── Makefile                  # Development commands
└── README.md                 # This file

Environment Variables

Variable	Description	Required
NOSANA_OLLAMA_URL	Nosana GPU endpoint URL	Yes
NOSANA_MODEL	Model name (default: gpt-oss:20b)	No
BRIGHTDATA_API_TOKEN	BrightData API authentication token	Yes
PORT	Backend server port (default: 8000)	No

License

ISC