firecrawl/rag-arena

Implements a competitive RAG evaluation framework where multiple retrieval strategies (vector store, graph-based, multi-query, contextual compression) compete to answer the same queries, with user votes updating Elo ratings stored in Supabase. The Next.js frontend interfaces with a Flask Python service that runs LangChain retrievers, while a dynamic routing layer selects which RAG method to use per query. Real-time leaderboards track retriever performance across voting metrics, enabling systematic comparison of different embedding and document chunking approaches.

216 stars. No commits in the last 6 months.