MUTHUKUMARAN-K-1/graphrag-inference-hackathon — reverse-engineered prompt

Build me a web app for a GraphRAG benchmark demo.

I want one question to run through three pipelines at the same time, LLM only, basic RAG, and GraphRAG, then show the answers side by side so it is obvious which one is faster, cheaper, and better. Please include a simple playground where I can type a question, paste my own API key in the UI, keep that key stored locally in the browser, and see response text, token counts, estimated cost, latency, and a few quality signals.

I also want a benchmarks page that can run a batch of science questions against the provided corpus and show summary metrics like F1, exact match, BERTScore, pass or fail from a judge model, and some clean charts for comparison. Add an explorer page too, so I can visually browse the knowledge graph and inspect related entities and links.

Please use the repo data and ingestion pieces that already exist, make the whole thing easy to run locally, and look up current docs online if you need to.