Atom

Perplexity-style AI search engine. Searches the web in real time, scrapes and ranks sources, streams a cited answer via SSE.

Next.js 16 · Express 5 · TypeScript · MongoDB · Redis

How It Works

Query → Optimize → Search (Serper) → Scrape (Cheerio/BullMQ) → Rank Sources → Stream LLM Answer → Validate → Store Memory

• Multi-provider LLM — NVIDIA NIM (GLM-5.2, Mistral Medium, Nemotron Ultra, MiniMax M3, DeepSeek V4), OpenAI GPT-4o-mini, Claude 3.5 Haiku, Gemini 2.5 Flash (native streaming for all)
• Source ranking — relevance, domain authority, freshness
• Vector memory — MongoDB Atlas Search for long-term recall across conversations (authenticated users only)
• Answer validation — post-generation hallucination check
• Search modes — web, news, academic

Setup

git clone https://github.com/SpiritOfKedar/atom.git && cd atom

# Server
cd server && cp .env.example .env && npm install && npm run dev

# Client (in another terminal)
cd client && cp .env.example .env && npm install && npm run dev

Fill in your .env files — at minimum you need one LLM key (NVIDIA_API_KEY / NVAPI_KEY, OPENAI_API_KEY, ANTHROPIC_API_KEY, or GEMINI_API_KEY). SERPER_API_KEY is required in production. MongoDB required, Redis recommended in production.

Production Deployment (Vercel + Railway/Render)

Vercel (client)

Variable	Description
NEXT_PUBLIC_API_URL	Public URL of the Express API (e.g. https://your-api.railway.app)
NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY	Clerk publishable key
CLERK_SECRET_KEY	Clerk secret key
SERPER_API_KEY	Serper API key for Discover news feed

In the Clerk dashboard, add your Vercel production (and preview) domains to allowed origins.

Railway / Render (API)

Variable	Description
NODE_ENV	production
PORT	Set by platform (often injected automatically)
CORS_ORIGIN	Your Vercel app URL (e.g. https://your-app.vercel.app)
MONGODB_URI	MongoDB Atlas connection string (TLS, with credentials)
REDIS_URL	Managed Redis URL (recommended for cache, BullMQ, guest quotas)
SERPER_API_KEY	Serper API key
CLERK_SECRET_KEY	Clerk secret key
At least one LLM key	NVIDIA_API_KEY, OPENAI_API_KEY, ANTHROPIC_API_KEY, or GEMINI_API_KEY

Build & start commands:

npm run build
npm start

Health probes:

• Liveness: GET /api/health
• Readiness: GET /api/health/ready (returns 503 if MongoDB is disconnected)

MongoDB Atlas vector index

Vector memory requires a MongoDB Atlas Search vector index on the memories collection:

1. Create an Atlas cluster with authentication and network access (allow Railway/Render egress IPs).
2. In Atlas Search, create a vector search index on collection memories:
   • Index name: vector_index
   • Field: embedding (vector, dimensions matching your embedding model — OpenAI text-embedding-3-small uses 1536)
3. Confirm the index name matches server/src/services/vector-store.service.ts.

Guest access

• Unauthenticated users get GUEST_MESSAGE_LIMIT free chat requests (default: 2), enforced server-side by IP.
• Set REQUIRE_AUTH_FOR_CHAT=true in production to require sign-in for all chat requests.
• Vector memory is stored only for authenticated users.

API

POST /api/chat

{
  "query": "What is quantum computing?",
  "searchType": "web | news | academic",
  "answerStyle": "concise | detailed | bullet-points",
  "modelProvider": "z-ai/glm-5.2 | mistralai/mistral-medium-3.5-128b | nvidia/nemotron-3-ultra-550b-a55b | minimaxai/minimax-m3 | deepseek-ai/deepseek-v4-pro | deepseek-ai/deepseek-v4-flash | openai | claude | gemini",
  "conversationId": "optional"
}

Returns SSE stream with events: status, sources, token, validation, followUps, conversationId, error.

Requires Authorization: Bearer <token> for saved conversations. Guests are rate-limited by IP.

Other endpoints

• GET /api/health — liveness check
• GET /api/health/ready — readiness check (MongoDB)
• GET /api/health/detailed — dependency status (authenticated)
• GET /api/conversations — list conversations (authed)
• GET /api/conversations/:id — get conversation (authed)
• DELETE /api/conversations/:id — delete conversation (authed)

Development scripts

Package	Command	Description
server	npm run dev	Start API with nodemon
server	npm run build	Compile TypeScript to dist/
server	npm start	Run compiled production server
server	npm test	Run unit tests
client	npm run dev	Start Next.js dev server
client	npm run build	Production build

License

MIT — see LICENSE.