Hargurjeet/hybrid-rag-nextjs

Hybrid RAG Research Assistant

A conversational research assistant that searches across 10,000 arXiv papers using a hybrid retrieval approach (BM25 + dense vector search), Cohere reranking, and Groq (Llama 3) for answer generation.

Features

• —Hybrid retrieval: Combines BM25 keyword search with dense semantic vectors for more relevant results
• —Cohere reranking: Re-scores candidates using a cross-encoder for precision
• —Multi-turn chat: Ask follow-up questions in the same conversation
• —Source documents: Each answer cites the specific papers it drew from
• —Evaluation tab: Run RAGAS faithfulness evaluation against the live pipeline

Stack

• —Backend: FastAPI + ChromaDB + Cohere + Groq (Llama 3.1 70B)
• —Frontend: Next.js 16, React 19, Tailwind CSS v4, shadcn/ui
• —Dataset: 10,000 arXiv papers (astro-ph, hep-th, hep-ph, quant-ph, gr-qc, cond-mat)