WAR — Global Conflict Monitor

A real-time global conflict monitoring dashboard that aggregates multi-lingual news sources, maps active conflict zones, and provides AI-powered intelligence analysis.

Live: war-pied-two.vercel.app

Credits

This project is inspired by and built upon NewFeeds by Maria Sukhareva (@ktoetotam).

Maria originally built a public monitoring website to track the war in Iran, aggregating diverse multi-lingual sources — Farsi, Hebrew, Arabic, Russian, Chinese — including Telegram channels and state-controlled media. Her philosophy: monitoring lies can reveal truth, and source diversity is essential during conflict.

Read her original write-up: I Built a Public Monitoring Website For The War in Iran — AI Realist, March 2026.

What We Added

Starting from the NewFeeds foundation, we extended the project with a full AI-powered intelligence layer using Retrieval-Augmented Generation (RAG) and a multi-agent system.

Interactive Conflict Map

• Real-time world map with clustered markers (Leaflet + MarkerCluster)
• Severity-based color coding (Low, Medium, High, Critical, Major)
• Clickable filter badges to toggle conflict severity levels
• Floating transparent header that overlays the map

AI Intelligence Chat (Multi-Agent RAG)

• Multi-agent RAG system powered by LlamaIndex and Groq
• 4 specialized agents: Intel, Bias Analysis, Fact-Check, Briefing
• Real-time streaming responses with markdown formatting
• Suggested questions for quick exploration

Bias Analyzer

• AI-driven media bias detection across sources
• Compares framing, tone, and emphasis across different outlets
• Highlights potential propaganda and state-controlled narratives

Fact Checker

• Cross-references claims against multiple sources
• Provides evidence-based verdicts with source citations

News Feed

• Aggregated news from diverse global sources
• Category and region tagging with pastel color system
• Relative timestamps and source attribution

Architecture

RAG Pipeline (Retrieval-Augmented Generation)

The core intelligence system uses RAG to ground AI responses in real, up-to-date news data rather than relying on the LLM's training data alone.

News Sources → Fetch & Translate → Embed → Store in pgvector
                                                    ↓
User Query → Embed Query → Semantic Search → Context → LLM → Response

How it works:

1. Ingestion (Cron Job): A scheduled cron job fetches articles from diverse multi-lingual news sources. Each article is:

   • Translated to English (if needed) and summarized using Groq LLM (llama-3.1-8b-instant)
   • Classified for conflict relevance — the LLM extracts attack type, severity, location
   • Geocoded via OpenStreetMap Nominatim for map placement
   • Embedded into a 1024-dimensional vector using a custom vLLM embedding model
   • Stored in Neon PostgreSQL with the pgvector extension

2. Retrieval: When a user asks a question, the query is embedded with the same model and a cosine similarity search (<=> operator) finds the most relevant articles in the database.

3. Generation: The retrieved articles are passed as context to the LLM, which generates a grounded, source-cited response.

Multi-Agent System

Built with LlamaIndex TS (LLMAgent), the system uses 4 specialized agents, each with access to different tool combinations:

Agent	Purpose	Tools
Intel Agent	Main chat — answers any geopolitical question	All 5 tools
Bias Agent	Compares how different sources frame the same event	search_news, get_recent_news, analyze_bias
Fact-Check Agent	Verifies claims against multiple sources	search_news, get_attacks, fact_check
Briefing Agent	Generates situational intelligence briefings	get_recent_news, get_attacks, search_news

Agent Tools:

Tool	Description	Technique
search_news	Semantic search over articles	pgvector cosine similarity
get_recent_news	Latest articles by region	Recency-based SQL query
get_attacks	Recent military/security incidents	Time-windowed attack data
analyze_bias	Compare source framing	LLM analysis over multi-source articles
fact_check	Verify claims with evidence	Cross-reference search + LLM verdict

Agents are powered by Groq (free tier) with two models:

• llama-3.1-8b-instant — fast model for bulk tasks (translation, classification, summarization)
• llama-4-scout-17b-16e-instruct — quality model for agent reasoning and chat responses

Streaming Architecture

Chat responses use a 2-phase streaming approach via Server-Sent Events (SSE):

1. Phase 1 — Agent Gathering: The LlamaIndex agent calls its tools (search, attacks, bias analysis) to gather intelligence. During this phase, the UI shows a status message ("Gathering intelligence...").

2. Phase 2 — Streamed Formatting: The raw agent output is passed to Groq's streaming API, which reformats it into clean markdown and streams it token-by-token to the client.

Client ←SSE← [sources] [status: "Gathering..."] [token][token][token]... [done]

Database Schema

PostgreSQL (Neon) with pgvector extension:

• sources — news source registry (name, URL, region, language, category)
• articles — fetched articles with embedding vector(1024) column for semantic search
• attacks — classified conflict events with lat/lon for map markers
• bias_analyses — cached bias analysis results
• fact_checks — cached fact-check verdicts with evidence

Rate Limiting

A custom queue-based rate limiter with exponential backoff handles Groq's free tier limits (RPM, RPD, TPM per model). All LLM calls are routed through this limiter to prevent 429 errors.

Tech Stack

• Framework: Next.js 15 + React 19 + TypeScript
• AI/LLM: LlamaIndex TS, Groq (Llama 3.1 / Llama 4 Scout)
• RAG: pgvector semantic search + custom embeddings (1024-dim)
• Database: Neon PostgreSQL + pgvector
• Embeddings: Custom vLLM embedding server
• Map: Leaflet + MarkerCluster (CDN)
• Styling: Bootstrap 5, custom dark theme with pastel accents
• Deployment: Vercel

Project Structure

src/
├── app/
│   ├── api/
│   │   ├── chat/          # AI chat endpoint (streaming SSE)
│   │   ├── bias/          # Bias analysis endpoint
│   │   ├── factcheck/     # Fact-checking endpoint
│   │   ├── news/          # News aggregation endpoint
│   │   └── cron/          # Scheduled news fetching
│   ├── globals.css        # Dark theme + pastel design system
│   ├── layout.tsx
│   └── page.tsx           # Main app with tab navigation
├── components/
│   ├── ConflictMap.tsx     # Interactive map with clustering
│   ├── ChatPanel.tsx       # AI chat with streaming markdown
│   ├── NewsFeed.tsx        # News aggregation feed
│   ├── BiasAnalyzer.tsx    # Media bias analysis
│   └── FactChecker.tsx     # Fact-checking interface
└── lib/
    ├── llamaindex/
    │   ├── config.ts       # LLM configuration (Groq)
    │   ├── agents.ts       # Multi-agent system (4 agents)
    │   └── tools.ts        # Agent tools (search, analyze, etc.)
    ├── services/           # Business logic services
    ├── db.ts               # PostgreSQL + pgvector
    ├── embeddings.ts       # Vector embeddings
    └── groq.ts             # Groq client + rate limiter

Getting Started

Prerequisites

• Node.js 18+
• PostgreSQL with pgvector extension (e.g., Neon)
• Groq API key
• (Optional) Custom embedding server

Environment Variables

Create a .env.local file:

DATABASE_URL=postgresql://...
GROQ_API_KEY=gsk_...
EMBEDDING_API_URL=        # Optional: custom embedding server
EMBEDDING_API_KEY=        # Optional
EMBEDDING_MODEL=          # Optional
NEWSDATA_API_KEY=         # For news fetching

Install & Run

npm install
npm run dev

Open http://localhost:3000.

License

MIT — same as the original NewFeeds project.