News Hound: The Only News Agent You'll Ever Need

Overview

One Search. Multiple Sources. One Final Answer. Zero Hallucinations.

This is a high-performance AI search agent that orchestrates multiple search engines to deliver synthesized, accurate answers. Think of it as a "Google AI Overviews" engine built for developers, optimized for speed and zero hallucination.

You can plug this agent into:

• RAG Retrieval Pipelines to get accurate summarized information from the web
• Real world agents that need fast accurate information about products and news from internet

Architecture

News_Hound/
├── Backend (Python/FastAPI)
│   ├── research_engine/ - Core AI research engine
│   │   ├── research_agent.py - Orchestrates search → synthesis pipeline
│   │   ├── local_llm.py - Llama-Swap & Gemini integration
│   │   ├── app.py - FastAPI server (port 2024)
│   │   ├── prompts.py - LLM prompt templates
│   │   └── graph.py - LangGraph state management
│   └── search_engines/ - Modular engine integrations
│       ├── brightdata.py - Social/video search
│       ├── tavily.py - Research transcripts
│       └── duckduckgo.py - Direct web search
│
├── Frontend (React/Vite/TypeScript)
│   ├── components/ - UI components
│   │   ├── InputForm.tsx - Query input
│   │   ├── ChatMessagesView.tsx - Results display
│   │   └── ActivityTimeline.tsx - Search progress
│   └── hooks/useStream.ts - WebSocket streaming
│
└── Infrastructure
    ├── Dockerfile + docker-compose.yml
    └── main.py - CLI entry point

Key Components

How It Works

1. Query Generation: LLM breaks down user question into 3 targeted sub-queries
2. Parallel Search: All 3 search engines execute simultaneously on each query
3. Snippet Synthesis: High-quality snippets extracted (not full-page scraping for speed)
4. Final Answer: Single synthesis pass produces cited answer with source links

Interface Preview

🔎 Intelligent Research Mode

🐾 Definitive Answers

Key Files

Dependencies

Backend:

• fastapi, uvicorn - Web framework
• langchain, langgraph - LLM orchestration
• google-genai - Gemini API client
• duckduckgo-search, requests - Search engines

Frontend:

• react, react-dom
• tailwindcss - Styling
• shadcn/ui - Component library

Quick Start

1. Install Dependencies

pip install -r requirements.txt
cd frontend && npm install

2. Configure API Keys

Copy the example environment file and fill in your keys:

cp .env.example .env

Open .env and add your credentials:

GEMINI_API_KEY=your_key
BRIGHTDATA_API_KEY=your_key
TAVILY_API_KEY=your_key

# DuckDuckGo doesn't need an API key!

# LLM Configuration
USE_GEMINI=False  # Set to True to use Gemini API instead of local LLM
GEMINI_MODEL=gemini-2.5-flash-lite  # Gemini model to use (when USE_GEMINI=True)
LOCAL_MODEL_PORT=8080  # use 8080 if running Llama-Swap, 11434 for Ollama
LOCAL_MODEL_NAME=qwen-opus  # Local model name (ignored if USE_GEMINI=True)
LOCAL_LLM_TIMEOUT=90  # Timeout for local LLM requests in seconds

3. Launch

# Start Backend (using uv for proper dependency management)
cd backend
uv run uvicorn research_engine.app:app --host 0.0.0.0 --port 2024 --reload

# Or use the Makefile command
make dev-backend

# Start Frontend (in a separate terminal)
cd frontend
npm run dev

Open http://localhost:5173/app/ to start searching.

Engine Optimization

Notable Features

• ⚡ 5x faster than traditional search via parallel execution
• 🔄 Zero-config - Just add API keys to .env
• 🎯 Cited answers - Every claim linked to source
• 📱 Real-time streaming - Watch research progress live
• 🐳 Docker-ready - Full containerization support

Advanced Usage

Each search engine can still be tested individually via CLI:

python search_engines/brightdata.py --search "query" --max 3
python search_engines/tavily.py --search "query" --max 3
python search_engines/duckduckgo.py --search "query" --max 3

Raw search result URLs and scoring metrics are always exported to search_results.tsv for manual review.

Output Files

After each research session, two output files are generated in the project root:

Command Line Usage

You can use the search endpoint directly from the command line without the UI:

Basic Search Examples

# Example 1: Who won the 2022 FIFA World Cup?
curl "http://localhost:2024/search?query=Who+won+the+World+Cup+in+2022"

# Example 2: What is the capital of France?
curl "http://localhost:2024/search?query=What+is+the+capital+of+France"

# Example 3: Latest news about artificial intelligence
curl "http://localhost:2024/search?query=Latest+news+about+artificial+intelligence"

Python CLI Tool

For a richer command-line experience with progress output, use the Python CLI:

# Basic search (uses settings from .env)
python news_hound.py --query "Who won the World Cup in 2022"

# With high effort
python news_hound.py --query "Who won the World Cup in 2022" --effort high

# With a specific model
python news_hound.py --query "Who won the World Cup in 2022" --model gemini-2.5-flash-lite

This tool reads all configuration from .env and outputs a formatted response with citations.

Advanced Options

# With effort levels
curl "http://localhost:2024/search?query=Who+won+the+World+Cup+in+2022&effort=low"
curl "http://localhost:2024/search?query=Who+won+the+World+Cup+in+2022&effort=high"

# Using a specific model
curl "http://localhost:2024/search?query=Who+won+the+World+Cup+in+2022&model=gemini-2.5-flash-lite"

Parameters:

• query (required): The search query
• effort (optional): low, medium, or high - affects depth of research
• model (optional): LLM model to use (default: gemini-2.5-flash-lite)

Response: Plain text with the synthesized answer and source citations.