Lumina
LiveLumina is a production-style AI search assistant built around a tool-first agent loop. Rather than forcing a web search on every query, the LLM autonomously decides when external information is needed, executing Tavily searches iteratively and returning streamed, cited answers with real-time sources.
Lumina highlights & features:
- Tool-first agent loop — the LLM decides when to search, eliminating forced pre-search on every query
- Live web search via Tavily with basic and advanced depth modes (search vs research)
- Streamed, cited answers with NDJSON streaming and inline [1][2] citations tied to real web sources
- Conversation memory retaining the last 20 messages with source markers stripped before LLM ingestion
- Multi-model routing swapping between GPT-4.1, Claude Sonnet 4, and Gemini 2.5 Flash via OpenRouter
- Auth & credit limits powered by Supabase JWT auth and server-side credit gate enforcement
- Follow-up suggestions with the model proposing relevant next questions at the end of each answer
A high-performance AI search engine combining tool calling, streamed citations, multi-model routing, and credit-gated session management.
Stack used
BunExpress 5React 19TypeScriptPostgreSQLPrismaTailwind CSS